A multi-attribute cloud resource demand prediction method based on drift perception and orthogonal frequency domain modeling

Abstract

The application relates to the computer technical field and provides a multi-attribute cloud resource demand prediction method based on drift perception and orthogonal frequency domain modeling. The method solves the problem that resource demand sequences in a cloud environment have strong coupling and non-stationary characteristics, so that existing prediction methods cannot simultaneously consider prediction accuracy and calculation efficiency. The scheme is as follows: normalizing a data set and embedding time characteristics; performing trend-residual decomposition to obtain residual of long-term trend and short-term fluctuation characteristics, and constructing a drift state representing resource demand distribution change; dynamically mixing the decomposed multi-attribute characteristics based on the drift state, realizing joint modeling of the time dimension and the attribute dimension; performing orthogonal frequency domain modeling on high-level feature representation, mapping the high-level feature representation to an orthogonal frequency domain space, and predicting frequency domain coefficients corresponding to a future prediction window; performing orthogonal inverse transformation on the frequency domain coefficients to obtain a time domain prediction result, and outputting corresponding multi-attribute cloud resource demand prediction values.

Description

Technical Field

[0001] This invention relates to the field of computer technology and provides a method for predicting multi-attribute cloud resource demand based on drift sensing and orthogonal frequency domain modeling. Background Technology

[0002] With the rapid development of cloud computing and big data technologies, the types of services carried by cloud platforms are constantly diversifying, and resource scheduling scenarios are becoming increasingly complex. To ensure the service quality of cloud computing platforms and improve resource utilization, it is usually necessary to predict the demands of various resources such as CPU, memory, disk I / O, and network bandwidth, and to perform elastic scaling, task scheduling, and capacity planning accordingly. Therefore, cloud resource demand forecasting has become one of the key technologies in cloud resource management. Most existing cloud resource demand forecasting methods are based on historical monitoring data to model future resource usage, which can reflect the changing patterns of resource demand to a certain extent. However, in real cloud environments, resource demand sequences are usually composed of multiple resource attributes, and different resource attributes have significant differences in physical meaning, numerical scale, and range of change. At the same time, affected by factors such as business load fluctuations, system configuration adjustments, task migration, and sudden requests, resource demand sequences often exhibit significant non-stationary characteristics. In addition, resource demand sequences usually contain both long-term trends and short-term fluctuation information, and their statistical characteristics change dynamically over time, increasing the difficulty of predictive modeling.

[0003] In summary, existing cloud resource demand forecasting methods have at least the following shortcomings:

[0004] 1) For multi-attribute resource demand sequences, the coupling relationship between different resource attributes is not fully utilized, making it difficult to simultaneously characterize the correlation in both the time dimension and the attribute dimension, resulting in insufficient ability to express complex resource demand change patterns.

[0005] 2) For non-stationary resource demand sequences, there is a lack of effective perception mechanisms for distribution drift and dynamic changes, making it difficult to adapt to scenarios where the statistical characteristics of resource demand change over time, thus affecting the accuracy and stability of prediction.

[0006] 3) For resource demand sequences that simultaneously contain both long-term trends and short-term fluctuations, existing methods often struggle to model the characteristics of different time scales, and error accumulation is likely to occur in multi-step prediction tasks.

[0007] 4) Although some existing methods can improve prediction accuracy, their model structures are relatively complex, their parameter scale is large, and their inference overhead is high. It is difficult to balance prediction performance and computational efficiency, which is not conducive to high-frequency prediction and online deployment in real cloud environments.

[0008] The aforementioned limitations of existing technologies restrict the application effectiveness of cloud resource demand forecasting models in complex and dynamic cloud environments, making it difficult to achieve high-precision, multi-step stable forecasting of multi-attribute resource demands while ensuring computational efficiency. Therefore, it is necessary to propose a new multi-attribute cloud resource demand forecasting method to improve the modeling capability of dynamic changes in resource demand, while balancing forecasting accuracy and computational efficiency. Summary of the Invention

[0009] The purpose of this invention is to address the problem that in complex cloud computing environments, the strong coupling and non-stationary characteristics of multi-attribute resource demand sequences make it difficult for existing cloud resource prediction methods to balance prediction accuracy and computational efficiency.

[0010] To achieve the above objectives, the present invention employs the following technical means:

[0011] This invention provides a multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling, comprising the following steps:

[0012] S1: Read cloud resource demand data and preprocess it, normalize the original sequence and embed time features to obtain a unified input representation;

[0013] S2: Perform trend-residual decomposition on the input sequence to separate long-term trend from short-term fluctuation information, and construct a drift state that characterizes the changes in resource demand distribution;

[0014] S3: Based on the drift state, perform dynamic feature mixing on the decomposed multi-attribute features to achieve joint modeling of the time dimension and the attribute dimension;

[0015] S4: Perform orthogonal frequency domain modeling on the fused high-level features to obtain the frequency domain coefficients corresponding to the future prediction window;

[0016] S5: The frequency domain prediction results are restored to the time domain through orthogonal inverse transformation, and the multi-attribute cloud resource demand prediction results are output.

[0017] In the above scheme, step S1 includes the following steps:

[0018] The detailed steps of data preprocessing in S1 of the above method are described below:

[0019] S11: Read historical resource demand data, assuming a multi-attribute resource demand time series. ,in Indicates the length of the historical observation window. Indicates the number of resources;

[0020] S12: Calculate the mean vector and standard deviation vector of the multi-attribute resource demand time series along the time dimension, respectively, using the following formulas:

[0021]

[0022]

[0023] in, Indicates the first The average value of each resource attribute within a historical observation window. Indicates the first The standard deviation of a resource attribute within a historical observation window To prevent numerically unstable smoothing terms;

[0024] S13: Based on the mean vector and standard deviation vector, the multi-attribute resource demand time series is normalized to obtain the normalized resource demand series, the calculation formula of which is:

[0025]

[0026] in Represents the normalized i-th The time step, the first The resource requirement value corresponding to each resource attribute;

[0027] S14: Extract the temporal features corresponding to the S13 sequence, denoted as follows: ,in The temporal feature dimension is represented; the temporal feature is projected onto a feature space consistent with the resource attributes through linear mapping to obtain the temporal embedding representation, the calculation formula of which is:

[0028]

[0029] in and Let represent the learnable weight parameters and bias parameters of the linear mapping of time features, respectively. Representation of time embedding;

[0030] S15: The time embedding representation is added element-wise to the normalized resource demand sequence to obtain a unified input representation, the calculation formula of which is:

[0031]

[0032] in, This indicates the input representation.

[0033] In the above scheme, to address the problem that the superposition of long-term trends and short-term fluctuations in resource demand sequences, and the dynamic drift of statistical characteristics over time, make it difficult to simultaneously ensure prediction accuracy and stability when directly modeling the original sequence, the following steps S21 to S28 are adopted. By performing trend-residual decomposition on the input representation and constructing drift states, the long-term evolution law and short-term disturbance information are structurally separated. At the same time, the distribution evolution characteristics such as load level migration, fluctuation intensity changes, and sudden behavior enhancement within the historical window are explicitly characterized, thereby reducing the interference of multi-timescale information aliasing and non-stationarity on feature extraction, improving the model's ability to identify the mechanism of resource demand changes, and its adaptive adjustment ability in the subsequent prediction process. Specifically, step S2 includes the following steps:

[0034] The detailed steps of decomposition and construction in S2 described above are as follows:

[0035] In real-world cloud resource demand forecasting scenarios, resource demand sequences typically superimpose both slowly evolving long-term trend components and short-term fluctuation components caused by sudden loads, task migrations, or scheduling disturbances. To reduce the interference of multi-timescale information aliasing on subsequent feature extraction, and to address the technical challenge of effectively distinguishing between long-term evolution patterns and short-term disturbance information when directly modeling the original input representation, further detailed steps such as S21 are employed, as follows:

[0036] S21: Input representation obtained in step S1 Low-pass filtering is performed to extract long-term trends; for the first... Each resource attribute, its trend component is represented as follows:

[0037]

[0038] in Indicates the first The time step, the first Trend values ​​corresponding to resource attributes Indicates the first The convolutional kernel weights corresponding to each resource attribute Indicates the size of the filtering window;

[0039] S22: Write the trend extraction process for each resource attribute in a holistic form:

[0040]

[0041] in Indicates the overall trend components, This indicates that the learnable convolution kernel parameters Defined trend extraction operator; after obtaining the trend components, the residual components are obtained by subtracting the trend components from the original input representation:

[0042]

[0043] in Residual components are used to characterize short-term fluctuations and high-frequency changes in resource demand sequences.

[0044] S23: The trend component With residual components The features are concatenated to form a joint representation:

[0045]

[0046] in Indicates joint expression, This represents a concatenation operation along the feature dimension;

[0047] In real-world cloud environments, resource demand distribution is often not static but rather gradually shifts with changes in business phases, resource scheduling strategies, and instantaneous access patterns. To address the technical challenge of existing methods effectively perceiving the overall horizontal migration and fluctuation characteristics of resource demand, and considering the need to explicitly characterize statistical differences between different historical phases, further detailed steps, such as S24, are employed as follows:

[0048] S24: Divide the historical observation window along the time dimension into two adjacent time sub-intervals, denoted as follows: and The mean vector and standard deviation vector for each of the two time sub-intervals are calculated using the following formulas:

[0049]

[0050]

[0051]

[0052]

[0053] in and These represent the lengths of the two consecutive time intervals;

[0054] S25: Perform difference operations on the statistical characteristics of the two time sub-intervals to characterize the changing trend of resource demand distribution over time, obtaining mean drift characteristics and fluctuation drift characteristics:

[0055]

[0056]

[0057] in Used to characterize the direction and magnitude of the migration of overall resource demand. Characterizing changes in the intensity of fluctuations in resource demand;

[0058] S26: Construct dynamic intensity features based on time differences. First, calculate the difference sequence between adjacent time steps:

[0059]

[0060] Further extract the average absolute change and the maximum absolute change, and their calculation formulas are as follows:

[0061]

[0062]

[0063] in Used to characterize the overall activity level of changes in resource demand. Used to characterize the extreme intensity of sudden changes in resource demand;

[0064] S27: Construct a ratio characteristic of high-frequency energy to low-frequency energy to reflect the proportion of high-frequency fluctuations relative to the overall energy. The calculation formula is as follows:

[0065]

[0066] Furthermore, the above drift statistics are globally summarized to obtain global scalar characteristics:

[0067]

[0068]

[0069]

[0070]

[0071] in and These represent the global variation magnitudes of the overall load level and the fluctuation intensity, respectively. This represents the overall intensity of sudden behavior within a resource demand sequence. This indicates the proportion of high-frequency changing components relative to the overall energy.

[0072] In the actual cloud resource demand forecasting process, the aforementioned mean drift characteristics, fluctuation drift characteristics, dynamic intensity characteristics, and energy ratio characteristics represent the evolution state of resource demand distribution from different perspectives. Based on the consideration of unifying and compressing multiple types of drift statistical information into a state representation that can be called by subsequent modules, in order to solve the technical problem that multi-source statistical features exist in a scattered manner and are difficult to directly participate in subsequent dynamic feature modulation, further detailed steps such as S28 are adopted, as follows:

[0073] S28: The mean drift feature, fluctuation drift feature, dynamic intensity feature, energy ratio feature, and global scalar feature are concatenated to form a high-dimensional statistical feature vector:

[0074]

[0075] The high-dimensional statistical feature vector is then compressed and fused using a nonlinear mapping function composed of a multilayer sensing mechanism to generate a drift state vector.

[0076]

[0077] in Represents the drift state vector. This represents a fully connected mapping function with a non-linear activation function.

[0078] In the above scheme, to address the technical problems of complex coupling relationships between cloud resource demand sequences in the time and attribute dimensions, high computational overhead of direct modeling, and difficulty in adapting to dynamic changes in resource demand distribution, the following steps S31 to S36 are adopted. Through time-slice partitioning, time-slice dimension mixing, channel-dimensional mixing, and drift-state driven feature modulation, joint modeling of multi-attribute resource demand sequences in the time and attribute dimensions is achieved, resulting in a high-level feature representation. Specifically, the detailed steps of dynamic feature mixing in S3 are described below:

[0079] S31: Joint feature representation obtained in step S2 Perform time-slicing based on the time dimension, assuming the length of a single time slice is... The sliding step size between adjacent time slices is Then the first Each time slice is represented as:

[0080]

[0081] in The number of time slices is represented by the following formula:

[0082]

[0083] when When, the time slice is divided into non-overlapping segments; when At that time, the time slice is divided into a sliding time slice with overlapping regions;

[0084] S32: Perform linear mapping on each time slice to obtain a fixed-dimensional time slice feature representation; let the time slice embedding dimension be... Then the first The embedding representation of each time slice is as follows:

[0085]

[0086] in Indicates the flattening operation. and These represent the weight and bias parameters of the time-slice mapping, respectively; the time-slice feature sequence is composed of all time slices. As input to the dynamic feature blending module;

[0087] In real-world resource demand forecasting scenarios, the distribution of resource demand dynamically evolves with changes in historical window states. If the subsequent feature mixing process adopts a fixed parameter mode, it is difficult to adaptively adjust to the current distribution state. Based on the consideration of enabling the feature interaction process to be guided by the drift state, and to solve the technical problem that traditional feature mixing methods lack the ability to adapt to distribution changes, further detailed steps such as S33 are adopted, as follows:

[0088] S33: Transfer the drift state vector obtained in step S2 Mapping to feature modulation parameters yields scaling and offset parameters, calculated using the following formulas:

[0089]

[0090] in Indicates the feature scaling parameter. Indicates the feature offset parameter. and Let represent the weight parameters and bias parameters of the linear mapping, respectively. Further, a feature linear modulation method is used to dynamically adjust the intermediate features, defined as:

[0091]

[0092] in Indicates the feature to be modulated. This represents element-wise multiplication;

[0093] S34: In the time slice dimension, after applying layer normalization and drift state modulation to the input time slice features, time slice mixing is performed through a one-dimensional convolution operator to obtain the time slice mixed branch output, the calculation formula of which is:

[0094]

[0095] in Presentation layer normalization operation, This demonstrates a one-dimensional convolution operator applied along the time slice dimension; the one-dimensional convolution adopts a depthwise separable convolution form, which is used to model the local dependencies between different time slices while reducing channel parameter coupling;

[0096] S35: In the feature channel dimension, after applying layer normalization and drift state modulation to the input time-slice features, channel mixing is performed through a multilayer perceptron to obtain the channel mixing branch output, the calculation formula of which is as follows:

[0097]

[0098] in This represents a channel mixing function consisting of two fully connected layers and a nonlinear activation function, used to characterize the nonlinear coupling relationship of different resource attributes in a high-dimensional feature space;

[0099] S36: The outputs of the time-slice mixing branch and the channel mixing branch are added element-wise and fused to obtain the output feature representation of the current dynamic feature mixing layer. The calculation formula is as follows:

[0100]

[0101] Multiple dynamic feature mixing layers are stacked to obtain a high-level feature representation:

[0102]

[0103] in This indicates the number of stacked layers in the dynamic feature blending layer. This represents the high-level feature representation of the output.

[0104] In the above scheme, step S3 includes the following steps:

[0105] The detailed steps of orthogonal frequency domain modeling in S3 described above are as follows:

[0106] S41: Globally converge the high-level feature representations obtained in step S3 along the time slice dimension to obtain the overall semantic representation. The calculation formula is as follows:

[0107]

[0108] S42: Construct an orthogonal frequency domain basis corresponding to the prediction window to characterize the different frequency components of the future resource demand sequence in the frequency domain space; let the prediction window length be... The corresponding orthogonal frequency domain basis matrix is ​​denoted as:

[0109]

[0110] S43: The overall semantic representation is mapped to the frequency domain coefficients corresponding to the future prediction window through linear mapping, resulting in the frequency domain coefficient matrix, the calculation formula of which is:

[0111]

[0112] in This represents the frequency domain coefficient matrix obtained from the prediction. and These represent the weight parameters and bias parameters of the frequency domain mapping, respectively.

[0113] S44: The frequency domain coefficient matrix is ​​used as the coefficient representation of the future prediction window on the orthogonal frequency domain basis, and passed to step S5 for orthogonal inverse transformation and prediction output.

[0114] In the above scheme, step S5 includes the following steps:

[0115] The detailed steps for outputting the prediction results in S5 using the above method are described below:

[0116] S51: Perform an orthogonal inverse transform on the frequency domain coefficient matrix obtained in step S4 to restore the frequency domain representation to the time domain prediction sequence. The calculation formula is as follows:

[0117]

[0118] in This represents the frequency domain coefficient matrix obtained in step S4. This represents the recovered time-domain predicted sequence;

[0119] S52: Using the statistics obtained during the normalization process in step S1, the time-domain prediction sequence is de-normalized to restore the original resource demand numerical scale. The calculation formula is as follows:

[0120]

[0121]

[0122]

[0123] in This represents the final prediction result after inverse normalization. and These represent the standard deviation and mean of the corresponding resource attributes, respectively. This indicates element-wise multiplication.

[0124] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0125] 1. Enhance input data representation capabilities: By normalizing and embedding time features into the multi-attribute cloud resource demand sequence, the differences in dimensionality and distribution between different resource attributes can be effectively alleviated, enhancing the stability of the input representation and providing a unified and effective feature foundation for subsequent prediction modeling.

[0126] 2. Enhance the ability to characterize long-term trends and short-term fluctuations: By using trend-residual decomposition, the long-term trend and short-term fluctuation information in the resource demand sequence are separated, reducing the mutual interference between features at different time scales, thereby improving the model's ability to model complex resource demand change patterns.

[0127] 3. Improve adaptability to non-stationary changes: By constructing drift states, the changes in resource demand distribution are dynamically represented, enabling the model to perceive changes in the statistical characteristics of resource demand sequences, thereby improving the predictive ability and robustness of resource demand changes in non-stationary scenarios.

[0128] 4. Achieve joint modeling of time and attribute dimensions: By using the dynamic feature mixing module to jointly model time slice features and channel features, it is possible to simultaneously explore the dependencies in the time dimension and the coupling relationships in the attribute dimension, thereby improving the overall accuracy of multi-attribute cloud resource demand prediction.

[0129] 5. Balancing prediction accuracy and computational efficiency: By introducing an orthogonal frequency domain modeling strategy, the system improves the stability and accuracy of multi-step predictions while maintaining high computational efficiency, making it more suitable for practical cloud resource scheduling and capacity planning scenarios.

[0130] 6. This invention creatively combines the technical features of "trend-residual decomposition" and "drift state modeling" (S21-S28). The key new effect of this combination is that it first separates long-term changes from short-term disturbances, and then characterizes distribution drift based on the decomposition results. This makes the drift state no longer a rough statistical representation of the original mixed sequence, but can more accurately distinguish between overall load migration, increased volatility, and amplified sudden behavior. This solves the problem that information from multiple time scales is mixed together in the original resource sequence, making it difficult to accurately identify distribution changes. Decomposition improves the identifiability of drift modeling, and drift modeling, in turn, allows the decomposition results to further serve subsequent prediction and adjustment.

[0131] 7. This invention creatively combines the technical features of "drift state," "feature linear modulation," and "dynamic feature mixing" (S33-S36). The key new effect of this combination is that it maps the drift state into modulation parameters, which directly affect the subsequent time-slice mixing and channel mixing processes, enabling the feature interaction method to adaptively adjust according to changes in the current resource demand distribution. This solves the problem that traditional feature mixing processes have fixed parameters and are difficult to adapt to changes in non-stationary resource demand. It transforms the originally static feature interaction process into a distribution-aware, driven dynamic interaction process.

[0132] In summary, this invention significantly improves the accuracy, stability, and computational efficiency of multi-attribute cloud resource demand forecasting through techniques such as normalization and time feature embedding, trend-residual decomposition, drift state modeling, dynamic feature mixing, and orthogonal frequency domain modeling. It can better adapt to complex and dynamic cloud computing environments and is of great significance for optimizing cloud resource management, resource scheduling, and capacity planning. Attached Figure Description

[0133] Figure 1 A multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling is provided for the implementation of this invention;

[0134] Figure 2 This is a flowchart of establishing a prediction model in an example of the present invention;

[0135] Figure 3 This is a structural diagram of the prediction model established in the example of the present invention;

[0136] Figure 4 This is a diagram of the dynamic feature mixing module of the model in the example of the present invention. Detailed Implementation

[0137] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0138] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0139] Example 1

[0140] This invention provides a multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling, comprising the following steps:

[0141] S1: Read cloud resource demand data and preprocess it, normalize the original sequence and embed time features to obtain a unified input representation;

[0142] S2: Perform trend-residual decomposition on the input sequence to separate long-term trend from short-term fluctuation information, and construct a drift state that characterizes the changes in resource demand distribution;

[0143] S3: Based on the drift state, perform dynamic feature mixing on the decomposed multi-attribute features to achieve joint modeling of the time dimension and the attribute dimension;

[0144] S4: Perform orthogonal frequency domain modeling on the fused high-level features to obtain the frequency domain coefficients corresponding to the future prediction window;

[0145] S5: The frequency domain prediction results are restored to the time domain through orthogonal inverse transformation, and the multi-attribute cloud resource demand prediction results are output.

[0146] In the above scheme, step S1 includes the following steps:

[0147] The detailed steps of data preprocessing in S1 of the above method are described below:

[0148] S11: Read historical resource demand data, assuming a multi-attribute resource demand time series. ,in Indicates the length of the historical observation window. Indicates the number of resources;

[0149] S12: Calculate the mean vector and standard deviation vector of the multi-attribute resource demand time series along the time dimension, respectively, using the following formulas:

[0150]

[0151]

[0152] in, Indicates the first The average value of each resource attribute within a historical observation window. Indicates the first The standard deviation of a resource attribute within a historical observation window To prevent numerically unstable smoothing terms;

[0153] S13: Based on the mean vector and standard deviation vector, the multi-attribute resource demand time series is normalized to obtain the normalized resource demand series, the calculation formula of which is:

[0154]

[0155] in Represents the normalized i-th The time step, the first The resource requirement value corresponding to each resource attribute;

[0156] S14: Extract the temporal features corresponding to the S13 sequence, denoted as follows: ,in The temporal feature dimension is represented; the temporal feature is projected onto a feature space consistent with the resource attributes through linear mapping to obtain the temporal embedding representation, the calculation formula of which is:

[0157]

[0158] in and Let represent the learnable weight parameters and bias parameters of the linear mapping of time features, respectively. Representation of time embedding;

[0159] S15: The time embedding representation is added element-wise to the normalized resource demand sequence to obtain a unified input representation, the calculation formula of which is:

[0160]

[0161] in, This indicates the input representation.

[0162] In the above scheme, to address the problem that the superposition of long-term trends and short-term fluctuations in resource demand sequences, and the dynamic drift of statistical characteristics over time, make it difficult to simultaneously ensure prediction accuracy and stability when directly modeling the original sequence, the following steps S21 to S28 are adopted. By performing trend-residual decomposition on the input representation and constructing drift states, the long-term evolution law and short-term disturbance information are structurally separated. At the same time, the distribution evolution characteristics such as load level migration, fluctuation intensity changes, and sudden behavior enhancement within the historical window are explicitly characterized, thereby reducing the interference of multi-timescale information aliasing and non-stationarity on feature extraction, improving the model's ability to identify the mechanism of resource demand changes, and its adaptive adjustment ability in the subsequent prediction process. Specifically, step S2 includes the following steps:

[0163] The detailed steps of decomposition and construction in S2 described above are as follows:

[0164] In real-world cloud resource demand forecasting scenarios, resource demand sequences typically superimpose both slowly evolving long-term trend components and short-term fluctuation components caused by sudden loads, task migrations, or scheduling disturbances. To reduce the interference of multi-timescale information aliasing on subsequent feature extraction, and to address the technical challenge of effectively distinguishing between long-term evolution patterns and short-term disturbance information when directly modeling the original input representation, further detailed steps such as S21 are employed, as follows:

[0165] S21: Input representation obtained in step S1 Low-pass filtering is performed to extract long-term trends; for the first... Each resource attribute, its trend component is represented as follows:

[0166]

[0167] in Indicates the first The time step, the first Trend values ​​corresponding to resource attributes Indicates the first The convolutional kernel weights corresponding to each resource attribute Indicates the size of the filtering window;

[0168] S22: Write the trend extraction process for each resource attribute in a holistic form:

[0169]

[0170] in Indicates the overall trend components, This indicates that the learnable convolution kernel parameters Defined trend extraction operator; after obtaining the trend components, the residual components are obtained by subtracting the trend components from the original input representation:

[0171]

[0172] in Residual components are used to characterize short-term fluctuations and high-frequency changes in resource demand sequences.

[0173] S23: The trend component With residual components The features are concatenated to form a joint representation:

[0174]

[0175] in Indicates joint expression, This represents a concatenation operation along the feature dimension;

[0176] In real-world cloud environments, resource demand distribution is often not static but rather gradually shifts with changes in business phases, resource scheduling strategies, and instantaneous access patterns. To address the technical challenge of existing methods effectively perceiving the overall horizontal migration and fluctuation characteristics of resource demand, and considering the need to explicitly characterize statistical differences between different historical phases, further detailed steps, such as S24, are employed as follows:

[0177] S24: Divide the historical observation window along the time dimension into two adjacent time sub-intervals, denoted as follows: and The mean vector and standard deviation vector for each of the two time sub-intervals are calculated using the following formulas:

[0178]

[0179]

[0180]

[0181]

[0182] in and These represent the lengths of the two consecutive time intervals;

[0183] S25: Perform difference operations on the statistical characteristics of the two time sub-intervals to characterize the changing trend of resource demand distribution over time, obtaining mean drift characteristics and fluctuation drift characteristics:

[0184]

[0185]

[0186] in Used to characterize the direction and magnitude of the migration of overall resource demand. Characterizing changes in the intensity of fluctuations in resource demand;

[0187] S26: Construct dynamic intensity features based on time differences. First, calculate the difference sequence between adjacent time steps:

[0188]

[0189] Further extract the average absolute change and the maximum absolute change, and their calculation formulas are as follows:

[0190]

[0191]

[0192] in Used to characterize the overall activity level of changes in resource demand. Used to characterize the extreme intensity of sudden changes in resource demand;

[0193] S27: Construct a ratio characteristic of high-frequency energy to low-frequency energy to reflect the proportion of high-frequency fluctuations relative to the overall energy. The calculation formula is as follows:

[0194]

[0195] Furthermore, the above drift statistics are globally summarized to obtain global scalar characteristics:

[0196]

[0197]

[0198]

[0199]

[0200] in and These represent the global variation magnitudes of the overall load level and the fluctuation intensity, respectively. This represents the overall intensity of sudden behavior within a resource demand sequence. This indicates the proportion of high-frequency changing components relative to the overall energy.

[0201] In the actual cloud resource demand forecasting process, the aforementioned mean drift characteristics, fluctuation drift characteristics, dynamic intensity characteristics, and energy ratio characteristics represent the evolution state of resource demand distribution from different perspectives. Based on the consideration of unifying and compressing multiple types of drift statistical information into a state representation that can be called by subsequent modules, in order to solve the technical problem that multi-source statistical features exist in a scattered manner and are difficult to directly participate in subsequent dynamic feature modulation, further detailed steps such as S28 are adopted, as follows:

[0202] S28: The mean drift feature, fluctuation drift feature, dynamic intensity feature, energy ratio feature, and global scalar feature are concatenated to form a high-dimensional statistical feature vector:

[0203]

[0204] The high-dimensional statistical feature vector is then compressed and fused using a nonlinear mapping function composed of a multilayer sensing mechanism to generate a drift state vector.

[0205]

[0206] in Represents the drift state vector. This represents a fully connected mapping function with a non-linear activation function.

[0207] In the above scheme, to address the technical problems of complex coupling relationships between cloud resource demand sequences in the time and attribute dimensions, high computational overhead of direct modeling, and difficulty in adapting to dynamic changes in resource demand distribution, the following steps S31 to S36 are adopted. Through time-slice partitioning, time-slice dimension mixing, channel-dimensional mixing, and drift-state driven feature modulation, joint modeling of multi-attribute resource demand sequences in the time and attribute dimensions is achieved, resulting in a high-level feature representation. Specifically, the detailed steps of dynamic feature mixing in S3 are described below:

[0208] S31: Joint feature representation obtained in step S2 Perform time-slicing based on the time dimension, assuming the length of a single time slice is... The sliding step size between adjacent time slices is Then the first Each time slice is represented as:

[0209]

[0210] in The number of time slices is represented by the following formula:

[0211]

[0212] when When, the time slice is divided into non-overlapping segments; when At that time, the time slice is divided into a sliding time slice with overlapping regions;

[0213] S32: Perform linear mapping on each time slice to obtain a fixed-dimensional time slice feature representation; let the time slice embedding dimension be... Then the first The embedding representation of each time slice is as follows:

[0214]

[0215] in Indicates the flattening operation. and These represent the weight and bias parameters of the time-slice mapping, respectively; the time-slice feature sequence is composed of all time slices. As input to the dynamic feature blending module;

[0216] In real-world resource demand forecasting scenarios, the distribution of resource demand dynamically evolves with changes in historical window states. If the subsequent feature mixing process adopts a fixed parameter mode, it is difficult to adaptively adjust to the current distribution state. Based on the consideration of enabling the feature interaction process to be guided by the drift state, and to solve the technical problem that traditional feature mixing methods lack the ability to adapt to distribution changes, further detailed steps such as S33 are adopted, as follows:

[0217] S33: Transfer the drift state vector obtained in step S2 Mapping to feature modulation parameters yields scaling and offset parameters, calculated using the following formulas:

[0218]

[0219] in Indicates the feature scaling parameter. Indicates the feature offset parameter. and Let represent the weight parameters and bias parameters of the linear mapping, respectively. Further, a feature linear modulation method is used to dynamically adjust the intermediate features, defined as:

[0220]

[0221] in Indicates the feature to be modulated. This represents element-wise multiplication;

[0222] S34: In the time slice dimension, after applying layer normalization and drift state modulation to the input time slice features, time slice mixing is performed through a one-dimensional convolution operator to obtain the time slice mixed branch output, the calculation formula of which is:

[0223]

[0224] in Presentation layer normalization operation, This demonstrates a one-dimensional convolution operator applied along the time slice dimension; the one-dimensional convolution adopts a depthwise separable convolution form, which is used to model the local dependencies between different time slices while reducing channel parameter coupling;

[0225] S35: In the feature channel dimension, after applying layer normalization and drift state modulation to the input time-slice features, channel mixing is performed through a multilayer perceptron to obtain the channel mixing branch output, the calculation formula of which is as follows:

[0226]

[0227] in This represents a channel mixing function consisting of two fully connected layers and a nonlinear activation function, used to characterize the nonlinear coupling relationship of different resource attributes in a high-dimensional feature space;

[0228] S36: The outputs of the time-slice mixing branch and the channel mixing branch are added element-wise and fused to obtain the output feature representation of the current dynamic feature mixing layer. The calculation formula is as follows:

[0229]

[0230] Multiple dynamic feature mixing layers are stacked to obtain a high-level feature representation:

[0231]

[0232] in This indicates the number of stacked layers in the dynamic feature blending layer. This represents the high-level feature representation of the output.

[0233] In the above scheme, step S3 includes the following steps:

[0234] The detailed steps of orthogonal frequency domain modeling in S3 described above are as follows:

[0235] S41: Globally converge the high-level feature representations obtained in step S3 along the time slice dimension to obtain the overall semantic representation. The calculation formula is as follows:

[0236]

[0237] S42: Construct an orthogonal frequency domain basis corresponding to the prediction window to characterize the different frequency components of the future resource demand sequence in the frequency domain space; let the prediction window length be... The corresponding orthogonal frequency domain basis matrix is ​​denoted as:

[0238]

[0239] S43: The overall semantic representation is mapped to the frequency domain coefficients corresponding to the future prediction window through linear mapping, resulting in the frequency domain coefficient matrix, the calculation formula of which is:

[0240]

[0241] in This represents the frequency domain coefficient matrix obtained from the prediction. and These represent the weight parameters and bias parameters of the frequency domain mapping, respectively.

[0242] S44: The frequency domain coefficient matrix is ​​used as the coefficient representation of the future prediction window on the orthogonal frequency domain basis, and passed to step S5 for orthogonal inverse transformation and prediction output.

[0243] In the above scheme, step S5 includes the following steps:

[0244] The detailed steps for outputting the prediction results in S5 using the above method are described below:

[0245] S51: Perform an orthogonal inverse transform on the frequency domain coefficient matrix obtained in step S4 to restore the frequency domain representation to the time domain prediction sequence. The calculation formula is as follows:

[0246]

[0247] in This represents the frequency domain coefficient matrix obtained in step S4. This represents the recovered time-domain predicted sequence;

[0248] S52: Using the statistics obtained during the normalization process in step S1, the time-domain prediction sequence is de-normalized to restore the original resource demand numerical scale. The calculation formula is as follows:

[0249]

[0250]

[0251]

[0252] in This represents the final prediction result after inverse normalization. and These represent the standard deviation and mean of the corresponding resource attributes, respectively. This indicates element-wise multiplication.

[0253] Experimental Example

[0254] 1. Description of Experimental Dataset

[0255] We used 7 consecutive days of resource usage data from 2000 servers in the Alibaba Cloud public cluster dataset (cluster-trace-v2018) for training and testing. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. The field descriptions of the dataset are shown in Table 1.

[0256] Table 1: Description of Dataset Fields

[0257] Field Name Field Description machine_id Record the machine number of the current data. time_stamp The timestamp of the recorded data, in seconds. cpu_util_percent The machine's CPU utilization. mem_util_percent The machine's memory usage. net_in Network ingress traffic for the machine. net_out The machine's network outbound traffic. disk_io_percent Disk usage of the machine.

[0258] 2. Test Environment Description

[0259] The test uses historical cloud resource demand data over 60 consecutive time steps to predict changes in cloud resource demand over the next 60 time steps. Model training uses the Adam optimizer with an initial learning rate of 0.0001, a batch size of 128, and a maximum training epoch of 50. The model has a feature embedding dimension of 64, two dynamic blending layers, a channel blending hidden layer dimension of 256, a time slice length of 10, and a time slice step size of 5. All experiments were conducted on an environment equipped with an NVIDIA RTX 4090 GPU, a 13th Gen Intel(R) Core(TM) i9-13900K CPU, and 64 GB of RAM, using PyTorch 1.13.1 and CUDA 11.6.

[0260] 3. Evaluation Indicators Explanation

[0261] To comprehensively evaluate the effectiveness of the cloud resource prediction model, the experiment used root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as evaluation metrics. The calculation formulas for the three metrics are shown below:

[0262]

[0263]

[0264]

[0265] in It is the actual value. It is a predicted value. This indicates the length of the time series. The smaller the values ​​of these three evaluation indicators, the smaller the error between the predicted and actual values, and the higher the accuracy of the prediction results.

[0266] 3. Test Instructions

[0267] This invention discloses a multi-attribute cloud resource demand prediction method based on drift perception and orthogonal frequency domain modeling. The core of this invention lies in improving the accuracy, stability, and computational efficiency of multi-attribute cloud resource demand prediction through a collaborative design of normalization and temporal feature embedding, trend-residual decomposition, drift state modeling, dynamic feature mixing, and orthogonal frequency domain modeling. Compared with traditional prediction methods, this invention can explicitly separate the long-term trend and short-term fluctuations in the resource demand sequence, and dynamically characterize the changes in resource demand distribution through drift states, enhancing the model's adaptability to non-stationary scenarios. Simultaneously, the joint modeling of the time and attribute dimensions improves the ability to express the coupling relationship of multi-attribute resource demands. Furthermore, the structured representation of the future prediction sequence through orthogonal frequency domain modeling enhances the stability of multi-step prediction. This verification test is mainly used to simulate practical application scenarios in cloud resource prediction tasks, evaluate the prediction performance, stability, and computational efficiency of this invention under multi-variable, multi-step prediction tasks, and verify its superior overall performance compared to existing mainstream prediction models.

[0268] I. Performance Testing and Result Analysis

[0269] The performance of the ODM model proposed in this invention in cloud resource prediction tasks was comprehensively evaluated. By comparing it with classic time series prediction models such as TCN, LSTM, Transformer, Informer, and PatchTST, the advantages of the proposed method in prediction accuracy were verified. Table 2 shows the comparison results of evaluation metrics for different models at a prediction step size of 60, including MAE, RMSE, and MAPE. Experimental results show that the proposed method outperforms other comparative models in all of the above evaluation metrics.

[0270] Table 2 Comparison of evaluation metrics for different models after 60 steps.

[0271] Model MAE RMSE MAPE TCN 0.192 0.402 1.349 LSTM 0.192 0.428 1.352 Transformer 0.190 0.403 1.332 Informer 0.184 0.401 1.332 PatchTST 0.177 1.321 0.369 ODM 0.175 1.321 0.359

[0272] II. Module Testing and Result Analysis

[0273] To verify the contribution of each module to the performance of the ODM model proposed in this invention, an ablation experiment was conducted. The performance of the model after removing different modules was compared with that of the complete model on three evaluation metrics: MAE, RMSE, and MAPE. The impact of each component module on prediction performance was analyzed. The experimental setup included the following six models:

[0274] ODM-Dct: A model with the orthogonal frequency domain modeling module removed;

[0275] ODM-Film: A model that removes drift-state-driven feature modulation modules;

[0276] ODM-Channel: A model that removes the channel feature blending module;

[0277] ODM-Token: A model that removes the time-slice feature blending module;

[0278] ODM-Decomp: A model that removes the trend-residual decomposition module;

[0279] ODM: A complete ODM model.

[0280] Table 3 shows the comparison of evaluation metrics for the models after removing different modules at a prediction step size of 60. Experimental results show that removing the orthogonal frequency domain modeling module weakens the model's ability to structurally represent future resource demand sequences and reduces the stability of multi-step predictions, indicating that orthogonal frequency domain modeling plays a crucial role in improving prediction accuracy. Removing the drift-state-driven feature modulation module reduces the model's adaptability to dynamic changes in resource demand distribution, demonstrating the significant role of the drift-aware mechanism in handling non-stationary resource demand sequences. Removing the channel feature mixing module weakens the model's ability to model the coupling relationships between different resource attributes, leading to a decrease in prediction performance, indicating that feature interaction in the channel dimension is an important factor in improving the accuracy of multi-attribute predictions. Removing the time-slice feature mixing module reduces the model's ability to characterize time-dimensional dependencies, indicating that time-slice feature mixing plays a key role in extracting temporal variation features. Removing the trend-residual decomposition module makes it difficult for the model to effectively separate long-term trends from short-term fluctuations, further reducing overall prediction performance, demonstrating the importance of trend-residual decomposition in improving the model's modeling ability. In summary, the orthogonal frequency domain modeling module, drift state modulation module, channel feature mixing module, time slice feature mixing module, and trend-residual decomposition module in the ODM model all play an important role in improving the performance of cloud resource demand prediction. The effective combination of these modules is the key to achieving high-precision and high-stability prediction in this invention. Removing any module will have varying degrees of impact on the model performance, thus verifying the necessity and effectiveness of the module design.

[0281] Table 3 Comparison of evaluation metrics for different modules after 60 steps.

[0282] Model MAE RMSE MAPE ODM-Dct 0.189 0.372 1.349 ODM-Film 0.184 0.369 1.344 ODM-Channel 0.183 0.365 1.353 ODM-Token 0.185 0.368 1.347 ODM-Decomp 0.181 0.368 1.334 ODM 0.175 0.359 1.321

[0283] III. Test Summary

[0284] This section presents performance and module tests of the invention. Through a series of comparative experiments, the effectiveness and advantages of the multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling are verified. Experimental results show that the invention can significantly improve prediction accuracy in cloud resource demand prediction tasks, and exhibits good stability and computational efficiency. It addresses the shortcomings of traditional models in long-term trend modeling, short-term fluctuation characterization, and adaptation to non-stationary distribution changes, thereby better meeting the practical needs of cloud resource scheduling, resource management, and capacity planning.

Claims

1. A multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling, characterized in that, Includes the following steps: S1: Read historical resource demand data as a dataset, preprocess the dataset, and perform normalization and time feature embedding to obtain a unified input representation; S2: Perform trend-residual decomposition on the input representation to obtain the trend representing the long-term change pattern and the residual representing the short-term fluctuation characteristics, and construct a drift state representing the change in resource demand distribution based on the input representation and the trend and residual. S3: Based on the drift state, perform dynamic feature mixing on the decomposed multi-attribute features to achieve joint modeling of the time dimension and the attribute dimension; S4: Perform orthogonal frequency domain modeling on the high-level feature representation, map the high-level feature representation to the orthogonal frequency domain space, and predict the frequency domain coefficients corresponding to the future prediction window; S5: Perform an orthogonal inverse transform on the frequency domain coefficients to obtain the time domain prediction result, and output the corresponding multi-attribute cloud resource demand prediction value.

2. The multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling according to claim 1, characterized in that, Step S1 includes the following steps: S11: Read historical resource demand data, assuming a multi-attribute resource demand time series. ,in Indicates the length of the historical observation window. Indicates the number of resources; S12: Calculate the mean vector and standard deviation vector of the multi-attribute resource demand time series along the time dimension, respectively, using the following formulas: in, Indicates the first The average value of each resource attribute within a historical observation window. Indicates the first The standard deviation of a resource attribute within a historical observation window To prevent numerically unstable smoothing terms; S13: Based on the mean vector and standard deviation vector, the multi-attribute resource demand time series is normalized to obtain the normalized resource demand series, the calculation formula of which is: in Represents the normalized i-th The time step, the first The resource requirement value corresponding to each resource attribute; S14: Extract the temporal features corresponding to the sequence obtained in S13. The temporal features are represented as follows: ,in The temporal feature dimension is represented; the temporal feature is projected onto a feature space consistent with the resource attributes through linear mapping to obtain the temporal embedding representation, the calculation formula of which is: in and Let represent the learnable weight parameters and bias parameters of the linear mapping of time features, respectively. Representation of time embedding; S15: The time embedding representation is added element-wise to the normalized resource demand sequence to obtain a unified input representation, the calculation formula of which is: in, This indicates the input representation.

3. The multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling according to claim 1, characterized in that, Step S2 includes the following steps: S21: Input representation obtained in step S1 Low-pass filtering is performed to extract long-term trends; for the first... Each resource attribute, its trend component is represented as follows: in Indicates the first The time step, the first Trend values ​​corresponding to resource attributes Indicates the first The convolutional kernel weights corresponding to each resource attribute Indicates the size of the filtering window; S22: Write the trend extraction process for each resource attribute in a holistic form: in Indicates the overall trend components, This indicates that the learnable convolution kernel parameters Defined trend extraction operator; after obtaining the trend components, the residual components are obtained by subtracting the trend components from the original input representation: in Residual components are used to characterize short-term fluctuations and high-frequency changes in resource demand sequences. S23: The trend component With residual components The features are concatenated to form a joint representation: in Indicates joint expression, This represents a concatenation operation along the feature dimension; S24: Divide the historical observation window along the time dimension into two adjacent time sub-intervals, denoted as follows: and The mean vector and standard deviation vector for each of the two time sub-intervals are calculated using the following formulas: in and These represent the lengths of the two consecutive time intervals. This refers to a smoothing term introduced to prevent numerical instability during the calculation of standard deviation. Indicates the sub-interval of the previous time. No. Resource demand feature vectors corresponding to each time step Indicates the next time interval The Middle Resource demand feature vectors corresponding to each time step; S25: Perform difference operations on the statistical characteristics of the two time sub-intervals to characterize the changing trend of resource demand distribution over time, obtaining mean drift characteristics and fluctuation drift characteristics: in Used to characterize the direction and magnitude of the migration of overall resource demand. Characterizing changes in the intensity of fluctuations in resource demand; S26: Construct dynamic intensity features based on time differences. First, calculate the difference sequence between adjacent time steps: in Indicates the first [number] observation within the historical observation window The feature vector of multi-attribute resource demand corresponding to each time step The expression indicates the first [number] observation within the historical observation window. Each time step corresponds to a multi-attribute resource demand feature vector; Further extract the average absolute change and the maximum absolute change, and their calculation formulas are as follows: in Used to characterize the overall activity level of changes in resource demand. The energy ratio characteristic is used to characterize the extreme sudden intensity of changes in resource demand; S27: Construct a ratio characteristic of high-frequency energy to low-frequency energy to reflect the proportion of high-frequency fluctuations relative to the overall energy. The calculation formula is as follows: Furthermore, the above drift statistics are globally summarized to obtain global scalar characteristics: in and These represent the global variation magnitudes of the overall load level and the fluctuation intensity, respectively. This represents the overall intensity of sudden behavior within a resource demand sequence. This indicates the proportion of high-frequency changing components relative to the overall energy. This represents the number of resource attribute channels, used to normalize the statistics on each attribute; S28: The mean shift feature Fluctuation and drift characteristics Dynamic intensity characteristics Energy ratio characteristics and global scalar features The vectors are concatenated to form a high-dimensional statistical feature vector. : The high-dimensional statistical feature vector is then compressed and fused using a nonlinear mapping function composed of a multilayer sensing mechanism to generate a drift state vector. in Represents the drift state vector. This represents a fully connected mapping function with a non-linear activation function.

4. The multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling according to claim 1, characterized in that, Step S3 includes the following steps: S31: Joint feature representation obtained in step S2 Perform time-slicing based on the time dimension, assuming the length of a single time slice is... The sliding step size between adjacent time slices is Then the first Each time slice is represented as: in The number of time slices is represented by the following formula: in This represents the length of a single time slice in the time dimension, that is, the number of time steps contained in each time slice. When, the time slice is divided into non-overlapping segments; when At that time, the time slice is divided into a sliding time slice with overlapping regions; S32: Perform linear mapping on each time slice to obtain a fixed-dimensional time slice feature representation; let the time slice embedding dimension be... Then the first The embedding representation of each time slice is as follows: in Indicates the flattening operation. and These represent the weight and bias parameters of the time-slice mapping, respectively; the time-slice feature sequence is composed of all time slices. As input to the dynamic feature blending module; S33: Transfer the drift state vector obtained in step S2 Mapping to feature modulation parameters yields scaling and offset parameters, calculated using the following formulas: in Indicates the feature scaling parameter. Indicates the feature offset parameter. and Let represent the weight parameters and bias parameters of the linear mapping, respectively. The intermediate hidden features are dynamically adjusted using feature linear modulation, and its definition is: in Indicates the feature to be modulated. This represents element-wise multiplication; S34: In the time slice dimension, after applying layer normalization and drift state modulation to the input time slice features, time slice mixing is performed through a one-dimensional convolution operator to obtain the time slice mixed branch output, the calculation formula of which is: in Presentation layer normalization operation, This demonstrates a one-dimensional convolution operator applied along the time slice dimension; the one-dimensional convolution adopts a depthwise separable convolution form, which is used to model the local dependencies between different time slices while reducing channel parameter coupling; S35: In the feature channel dimension, after applying layer normalization and drift state modulation to the input time-slice features, channel mixing is performed through a multilayer perceptron to obtain the channel mixing branch output, the calculation formula of which is as follows: in This represents a channel mixing function consisting of two fully connected layers and a nonlinear activation function, used to characterize the nonlinear coupling relationship of different resource attributes in a high-dimensional feature space; S36: The outputs of the time-slice mixing branch and the channel mixing branch are added element-wise and fused to obtain the output feature representation of the current dynamic feature mixing layer. The calculation formula is as follows: Multiple dynamic feature mixing layers are stacked to obtain a high-level feature representation: in This indicates the number of stacked layers in the dynamic feature blending layer. This represents the high-level feature representation of the output.

5. The multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling according to claim 1, characterized in that, Step S4 includes the following steps: S41: Globally converge the high-level feature representations obtained in step S3 along the time slice dimension to obtain the overall semantic representation. The calculation formula is as follows: in This indicates a global pooling aggregation operation performed on the high-level feature representations along the time slice dimension. S42: Construct an orthogonal frequency domain basis corresponding to the prediction window to characterize the different frequency components of the future resource demand sequence in the frequency domain space; let the prediction window length be... The corresponding orthogonal frequency domain basis matrix is ​​denoted as: S43: The overall semantic representation is mapped to the frequency domain coefficients corresponding to the future prediction window through linear mapping, resulting in the frequency domain coefficient matrix. The calculation formula is as follows: in This represents the frequency domain coefficient matrix obtained from the prediction. This indicates a tensor reshaping operation. and These represent the weight parameters and bias parameters of the frequency domain mapping, respectively; S44: The frequency domain coefficient matrix The coefficients of the future prediction window are represented on the orthogonal frequency domain basis and passed to step S5 for inverse orthogonal transformation and prediction output.

6. The multi-attribute cloud resource demand prediction method based on drift sensing and orthogonal frequency domain modeling according to claim 1, characterized in that, Step S5 includes the following steps: S51: Perform an orthogonal inverse transform on the frequency domain coefficient matrix obtained in step S4 to restore the frequency domain representation to the time domain prediction sequence. The calculation formula is as follows: in This represents the frequency domain coefficient matrix obtained in step S4. This represents the orthogonal discrete cosine inverse transform operation. This represents the recovered time-domain predicted sequence; S52: Using the mean vector and standard deviation vector obtained during the normalization process in step S1, the time-domain prediction sequence is denormalized to restore the original numerical scale of resource demand. The calculation formula is as follows: in This represents the final prediction result after inverse normalization. and These represent the standard deviation vector and mean vector corresponding to each resource attribute calculated in step S12, respectively. This indicates element-wise multiplication.

Images