Control method and device of computing power demand prediction system, and storage medium

Abstract

The invention discloses a control method and device of a computing power demand prediction system and a storage medium, and relates to the technical field of system resource management, and the method comprises the steps: splicing input condition features and random Gaussian noise, and carrying out the fusion of the condition features and random Gaussian noise, and obtaining a high-dimensional hidden space representation; performing sequence decoding on the high-dimensional hidden space representation, arranging according to a decoding sequence, and determining a resource demand prediction sequence in a future time period; based on the resource demand prediction sequence and a real load sequence, calculating weighted loss according to a preset quantile, and outputting a composite prediction error value; extracting long-range features of the resource demand prediction sequence and the real load sequence, and obtaining a scalar score through feature scaling processing; and performing weighted fusion on the composite prediction error value and the scalar score to construct a joint optimization target so as to alternately update a generator and a discriminator and iteratively generate a prediction generator model. The problem of low resource utilization rate is solved, and the server utilization rate is improved.

Description

Technical Field

[0001] This application relates to the field of system resource management technology, and in particular to a control method, device and storage medium for a computing power demand prediction system. Background Technology

[0002] In high-density data processing in data centers, the prediction of computing power demand directly impacts overall resource utilization efficiency and equipment energy consumption. Related technologies often employ an over-redundancy deployment strategy in predicting load peaks and troughs, reserving a large amount of physical resources in advance to form a buffer layer for sudden peaks. However, excessive resource reservation leads to high server idle rates, resulting in low resource utilization.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a control method, device, and storage medium for a computing power demand prediction system, aiming to solve the technical problem of high server idle rate caused by excessive reserved resources, which in turn leads to low resource utilization.

[0005] To achieve the above objectives, this application proposes a control method, device, and storage medium for a computing power demand prediction system, wherein the method includes: The conditional features of the input are concatenated with random Gaussian noise and fused to obtain a high-dimensional latent space representation; The high-dimensional latent space representation is decoded sequentially, and the resource demand prediction sequence for future time periods is determined according to the decoding order. Based on the resource demand prediction sequence and the actual load sequence, a weighted loss is calculated according to a preset quantile, and a composite prediction error value is output. Long-range features of the resource demand prediction sequence and the actual load sequence are extracted, and scalar scores are obtained through feature scalarization. The composite prediction error value and the scalar score are weighted and fused to construct a joint optimization objective, which alternately updates the generator and discriminator to iteratively generate the prediction generator model.

[0006] In one embodiment, resource indicator data is generated by encapsulating the resource data captured by the monitoring probe according to a preset period and the corresponding business tag vector of the resource data. Preprocess the resource indicator data within each preset period according to preset processing rules to generate a target sequence stream; The features of the target sequence stream are extracted, and the obtained latent vector is concatenated with the sine and cosine periodic codes to generate a conditional feature vector.

[0007] In one embodiment, the target sequence stream is segmented by a sliding window of a preset fixed length to generate continuous time series segments; The sequence features of the time series segment are extracted by convolution calculation and gated loop operation, and the latent vector representing the sequence features is output. Based on the timestamp information corresponding to the target sequence stream, the daily and weekly periodic components are calculated and concatenated by sine and cosine function mapping to generate the sine and cosine periodic code. The conditional feature vector is generated by concatenating the sine and cosine periodic codes with the latent vector according to a preset vector concatenation method.

[0008] In one embodiment, based on the dependencies of the modeling features, the high-dimensional latent space representation is sequentially subjected to residual connections, layer normalization, and feedforward network transformation to output a primary-level feature representation; The initial feature representation is processed iteratively to refine the features and capture multi-scale time dependencies, outputting an intermediate-level feature representation. The intermediate feature representation is mapped to the target dimension based on a linear projection layer, and the resource demand prediction sequence is generated by decoding according to the numerical constraint range.

[0009] In one embodiment, based on the resource demand prediction sequence and the actual load sequence, the residual between the predicted value and the actual value is calculated point by point according to a preset quantile to generate a quantile residual sequence; By identifying and suppressing abnormal residual values ​​in the quantile residual sequence through a preset threshold, a robust residual metric is output. The composite prediction error value is generated by weighting and fusing the robust residual metric according to the quantile weight coefficient.

[0010] In one embodiment, the resource demand prediction sequence after data normalization and the actual load sequence are recombined into a time-series tensor format to generate a standardized input tensor. By calculating the causal convolution of the input tensor and superimposing a spectral normalization constraint weight matrix on the causal convolution, a multidimensional high-level feature tensor is generated. Based on the global average pooling operation on the multidimensional high-level feature tensor, and by mapping the obtained long-range features to scalar scores through a fully connected layer, the scalar scores representing the authenticity of the sequence are generated.

[0011] In one embodiment, the composite prediction error value and the scalar score are weighted and summed according to preset weight coefficients to generate a joint optimization loss value; Based on the input of the joint optimization loss value, the parameter gradients of the generator and discriminator are calculated alternately at a preset update frequency; The parameter gradients are processed by gradient synchronization to drive model parameter updates and output the trained prediction generator model.

[0012] In one embodiment, a multidimensional load sequence is input to the prediction generator model, a decoder is invoked to perform parallel computation, the original prediction tensor for the future time period is output, and a preset quantile prediction curve is generated through the quantile calculation kernel. Based on the predicted curve of the first preset quantile and the actual load sequence, the residual calculation kernel is called to generate a real-time error sequence, and the prediction generator model is calibrated based on the real-time error sequence. Based on the calibrated prediction generator model, the prediction curve of the second preset quantile is updated to generate flexible scaling decisions.

[0013] In addition, to achieve the above objectives, this application also proposes a computing power demand prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the control method of the computing power demand prediction system as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the control method of the computing power demand prediction system as described above.

[0015] This application provides a control method for a computing power demand prediction system. The method involves first concatenating conditional features with random Gaussian noise and then fusing them through affine transformation to generate a high-dimensional latent space representation with a unified dimension. Subsequently, a multi-layer self-attention mechanism and a feedforward neural network are used to sequentially decode this latent space representation, generating a resource demand prediction sequence for future time periods. Next, based on the predicted sequence and the actual load sequence, a weighted loss is calculated using preset quantiles to output a composite prediction error value. Simultaneously, a scalar score representing the sequence's authenticity is calculated through multi-layer spectral normalization and dilated convolution. Finally, the composite prediction error value and the scalar score are weighted and fused with a gradient penalty mechanism to construct a joint optimization objective, driving the generator and discriminator to update alternately, ultimately generating a high-performance prediction generator model. This scheme effectively addresses the shortcomings of traditional prediction models in training stability, multi-scale feature capture, and abnormal fluctuation perception, achieving synergistic optimization of high-precision resource demand prediction and system stability, significantly improving the robustness of prediction results to extreme load changes and the reliability of scheduling decisions.

[0016] In summary, this application solves the problem of low resource utilization by combining convolution, gated recursive coding and generative adversarial learning, achieves high-precision computing power demand prediction, improves server utilization and reduces data center energy consumption. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the control method for the computing power demand prediction system of this application. Figure 2 This is a schematic diagram of the main process of this application; Figure 3 This is a flowchart illustrating the third embodiment of the control method for the computing power demand prediction system of this application. Figure 4 This is the overall flowchart of this application; Figure 5 This is a schematic diagram of the computing power demand prediction device of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] In predicting load peaks and troughs, related technologies often employ an over-redundant deployment strategy, reserving a large amount of physical resources in advance to form a buffer layer for sudden peaks. However, excessive resource reservation leads to high server idle rates, resulting in low resource utilization.

[0023] This application provides a solution: First, the input conditional features and random Gaussian noise are concatenated and fused to obtain a high-dimensional latent space representation. Then, the high-dimensional latent space representation is sequentially decoded, and the resource demand prediction sequence for future time periods is determined according to the decoding order. Next, based on the resource demand prediction sequence and the actual load sequence, a weighted loss is calculated according to a preset quantile, and a composite prediction error value is output. Then, long-range features of the resource demand prediction sequence and the actual load sequence are extracted, and a scalar score is obtained through feature scalarization. Finally, the composite prediction error value and the scalar score are weighted and fused to construct a joint optimization objective, which alternately updates the generator and discriminator, iteratively generating a prediction generator model.

[0024] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or computing power demand prediction device capable of performing the above functions. The following description uses a computing power demand prediction device as an example to illustrate this embodiment and the subsequent embodiments.

[0025] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0026] This application provides a control method for a computing power demand prediction system, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the control method for the computing power demand prediction system of this application.

[0027] In this embodiment, the control method of the computing power demand prediction system includes steps S10 to S50: Step S10: Concatenate the input conditional features with random Gaussian noise and fuse them to obtain a high-dimensional latent space representation.

[0028] In this embodiment, conditional features refer to feature vectors extracted from historical data that contain temporal patterns and periodic information. Random Gaussian noise refers to a set of values ​​randomly sampled from a normal distribution with a mean of zero and a variance of one. Concatenation refers to joining two vectors in a dimensional direction to form a longer vector. High-dimensional latent space representation refers to the numerical representation in a feature space with higher dimensions that cannot be directly understood.

[0029] As an alternative implementation, a new composite vector is formed by concatenating two conditional feature vectors from memory and randomly generated noise vectors using a hardware-level splicing circuit. This composite vector is then input into an affine transformation core for matrix multiplication and vector addition. The result is then passed through a linear rectified activation function to output a novel, uniformly dimensional high-dimensional latent space representation vector. This affine transformation core consists of a weight matrix and a bias vector. This method offers extremely high processing speed, high energy efficiency, and very low prediction latency, meeting the stringent real-time requirements of demanding scenarios.

[0030] As an alternative implementation, this method leverages a general-purpose graphics processing unit (GPU) and software library collaboration. Conditional eigenvectors and random noise vectors are loaded from host memory into the GPU's video memory. A highly optimized matrix operation library is invoked to perform vector concatenation operations in parallel on a streaming multiprocessor, generating a temporary merged vector. Another dedicated matrix multiplication kernel function is then called, multiplying the merged vector by a predefined weight matrix and adding a bias vector to the product. The final result is the high-dimensional latent space representation. This method offers excellent flexibility and is suitable for research and development phases where algorithms require frequent iteration and optimization, or for general-purpose applications where prediction latency requirements are not critical.

[0031] Step S20: Sequence decoding of the high-dimensional latent space representation, and arranging according to the decoding order to determine the resource demand prediction sequence for future time periods.

[0032] In this embodiment, sequence decoding refers to the process of generating the output sequence step by step according to the time step sequence. The resource demand prediction sequence refers to the numerical sequence of resource demand at consecutive future time points generated by the model.

[0033] As an optional implementation, a high-dimensional latent space representation vector is received from the memory controller and loaded as the initial hidden state into the on-chip cache. The decoding process proceeds stepwise in an autoregressive manner. At each time step, the attention computation unit first calculates the association weights between the current hidden state and all historical hidden states, generating a weighted context vector. This vector is then fed into a feedforward neural network unit, which performs two linear transformations and inserts a nonlinear activation function between them, outputting the predicted value for this time step. This predicted value is written back to memory as an element of the final output sequence and is also fed back to the next time step as part of the input. This process is repeated until a complete target-length sequence is generated. Here, the feedforward neural network is a fully connected network that includes linear transformations and nonlinear activation functions. This method boasts extremely high computational efficiency and low power consumption, meeting the stringent requirements of applications that are highly sensitive to latency.

[0034] As an alternative implementation, the high-dimensional latent space representation is copied to GPU memory and processed collaboratively by multiple streaming multiprocessors. Each streaming multiprocessor is responsible for processing one or more time steps in the sequence. In each layer, a wrapped self-attention kernel function is invoked. This function computes the query, key, and value vectors in parallel and performs a scaled dot product operation to obtain attention weights and output a context vector. The context vector is then input into the feedforward network kernel function, undergoing a nonlinear transformation through a combination of matrix multiplication and activation functions. The results from each time step are then combined into a complete prediction sequence. This method offers excellent flexibility and programmability, allowing for rapid model deployment and updates leveraging a mature software ecosystem.

[0035] Step S30: Based on the resource demand prediction sequence and the actual load sequence, calculate the weighted loss according to the preset quantile and output the composite prediction error value.

[0036] In this embodiment, the true load sequence refers to the sequence formed by arranging the resource usage values ​​corresponding to the same time period, which are actually collected by the monitoring system, in chronological order. Preset quantiles refer to pre-defined probability quantiles used to evaluate the error of the predicted value at different confidence levels. Weighted loss refers to the error value calculated by assigning different weights to the importance of the quantiles. The composite prediction error value refers to a scalar value reflecting the overall prediction deviation, calculated by combining the losses from multiple quantiles.

[0037] As an optional implementation, the resource demand prediction sequence and the actual load sequence are loaded into the memory buffer of the error calculation unit, and the prediction error for each quantile channel is calculated independently. The calculation process involves obtaining the residual between the predicted and actual values ​​at each time point, then asymmetrically weighting the positive and negative residuals according to the weight coefficients corresponding to that quantile, and finally averaging the weighted results across all time points to obtain the loss value for that quantile. The outputs of all quantile channels are fed into a weighted fusion module, where they are linearly combined according to preset importance weights to generate a comprehensive composite prediction error value. This method is computationally efficient and has low latency, making it suitable for scenarios with high training speed requirements.

[0038] As an alternative implementation, the predicted and actual sequences are loaded from main memory into the processor's cache. First, the residual at each position in the sequence is calculated. Then, based on the current quantile value, negative residuals are multiplied by the quantile weight, and positive residuals are multiplied by a factor of one minus the quantile weight, to obtain the weighted loss for each point. The average of all weighted losses is taken as the loss for that quantile. These loss values ​​are then weighted and summed according to preset fusion weights to generate a composite prediction error value. This method is flexible, low-cost, and suitable for experimental environments in the research or development phase where frequent adjustments to quantile settings are required.

[0039] Step S40: Extract long-range features from the resource demand prediction sequence and the actual load sequence, and obtain a scalar score through feature scalarization processing.

[0040] In this embodiment, spectral normalization refers to a mathematical method that stabilizes the training process by constraining the spectral norm of the weight matrix. Dilated convolution is a special convolutional method that expands the receptive field by inserting spaces between kernel elements. Scalar scoring refers to the final output as a single numerical score.

[0041] As an optional implementation, a resource demand prediction sequence and a real load sequence are loaded into the chip's input buffer. Based on these sequences, a dilated convolutional unit expands the receptive field by introducing a specific interval parameter during convolution calculation, extracting long-range dependency features and generating a feature matrix. A spectral normalization unit then performs singular value decomposition on the feature matrix output from this layer to obtain the maximum singular value. The feature matrix is ​​then scaled based on this maximum singular value. After multiple alternating processing steps, the final generated feature tensor is compressed into a vector form by a global pooling layer, and then reduced to a single scalar value by a fully connected layer, serving as the scalar score output. This method provides extremely high-speed score output, suitable for online discrimination scenarios with extremely high real-time requirements.

[0042] As an alternative implementation, the computation process is implemented through a software library. Based on the resource demand prediction sequence and the actual load sequence, a dilated convolution kernel function is called for processing. This dilated convolution kernel function expands the receptive field by inserting zero values ​​between the convolution kernel elements. After each layer of convolution computation is completed, a spectral normalization function is called to iteratively calculate the spectral norm of the weight matrix and apply normalization constraints. After multi-layer processing, global average pooling is used to convert the feature maps into vectors, and finally, a scalar score output is obtained through matrix-vector multiplication. This method is flexible, has low development costs, and allows for easy adjustment of the number of network layers, convolution parameters, and normalization methods through software updates, making it compatible with various model variants.

[0043] Step S50: The composite prediction error value and the scalar score are weighted and fused to construct a joint optimization objective, which is used to alternately update the generator and discriminator to iteratively generate the prediction generator model.

[0044] In this embodiment, weighted fusion refers to the mathematical operation of combining multiple values ​​into a single value according to preset weights. Gradient penalty mechanism refers to a regularization method that enhances model training stability by adding gradient norm constraints. Joint optimization objective refers to a unified optimization objective function formed by combining multiple loss functions. Generator refers to a neural network model used to generate predicted sequences. Discriminator refers to a neural network model used to evaluate the authenticity of sequences. Alternating update refers to a training strategy in which the generator and discriminator update their parameters in turn. Predictive generator model refers to the finally trained generator neural network capable of outputting predicted sequences.

[0045] As an optional implementation, the composite prediction error value and scalar score are input into the processing unit. The weighted fusion unit first reads preset weight coefficients from the weight memory, and then performs a weighted summation calculation on the composite prediction error value and the scalar score to obtain the base loss value. Simultaneously, the gradient penalty module calculates the gradient norm of the discriminator for the randomly interpolated samples using forward automatic differentiation, and multiplies the squared difference between this norm and the target value by a penalty coefficient to obtain a penalty term. The base loss value and the penalty term are fed into an adder to merge into a joint optimization target value, which is output to the training control system. The system calculates the gradient based on this value and updates the parameter weights of the generator and discriminator according to an alternating update strategy. After multiple iterations, the trained prediction generator model parameters are output. This method achieves fast and stable model training and is suitable for production environments with high training efficiency requirements.

[0046] As an alternative implementation, a weighted base loss value is obtained by calling linear algebra library functions to weight and fuse the input composite prediction error value and scalar score. An automatic differentiation framework is then used to calculate the gradient penalty term. This is achieved by constructing random interpolation samples and calculating the gradient norm of the discriminator for those samples. The squared difference between this norm and a set value is then calculated and multiplied by the penalty coefficient. Finally, the base loss value and the penalty term are added to obtain the joint optimization objective value. Based on this value, the parameter update amount is calculated, and the generator and discriminator parameters are updated separately using an alternating update strategy. After multiple iterations, a prediction generator model is obtained. This method is suitable for experimental scenarios where frequent adjustments to the optimization objective and hyperparameters are required during the research phase.

[0047] For example, multidimensional conditional features containing historical resource utilization and business tags are collected from a physical server cluster. These features are concatenated with random Gaussian noise vectors sampled from a standard normal distribution. Feature fusion is achieved through linear weighting and bias processing of an affine transformation layer, generating a high-dimensional latent space representation containing underlying patterns. This representation is input to a generator with a six-layer encoder-decoder architecture, which captures temporal dependencies through a multi-head self-attention mechanism and performs a nonlinear transformation via a feedforward neural network to output a predicted sequence of CPU / GPU / memory resource demands for the next 60 minutes. Simultaneously, this predicted sequence and the load sequence collected from actual monitoring are input into a discriminator with a five-layer temporal convolutional network architecture. Each layer is subject to spectral normalization constraints and performs convolution calculations with increasing dilation rates, ultimately outputting a scalar score representing the authenticity of the sequence. The system calculates Huber loss at 50% / 90% / 99% quantiles based on the predicted and true sequences, generates a weighted composite prediction error value, and fuses this error value with the scalar score through weight coefficients. A Wasserstein gradient penalty loss term based on gradient penalty at random interpolation points is added to form a joint optimization objective that drives the generator and discriminator to update alternately at a 5:1 frequency. After distributed training, the final prediction generator model is generated. After deployment, the model outputs quantile prediction curves in real time to drive flexible scaling decisions.

[0048] Furthermore, referring to Figure 2 , Figure 2 This is a schematic diagram of the main process of this application. In the scenario of predicting computing power requirements using generative adversarial networks, the system first concatenates the conditional feature input with random Gaussian noise and generates a high-dimensional latent space representation through affine fusion. Then, the Transformer Decoder performs sequence decoding on this representation to generate a resource demand prediction sequence for future time periods. Based on this prediction sequence and the real load sequence, the Huber-Quantile composite loss is calculated according to preset quantiles, and the composite prediction error value is output. At the same time, the discriminator extracts the long-range features of the two sequences and obtains the true / false score through feature scalarization. Finally, the composite prediction error value and the true / false score are weighted and fused with Wasserstein-GP loss to construct a joint optimization objective. The generator and discriminator are trained in an alternating update manner to iteratively generate the prediction generator model.

[0049] By using a generative adversarial network framework, the problem of inaccurate prediction of burst traffic and extreme values ​​in traditional time series prediction models is solved, thereby improving the accuracy of resource prediction and scheduling efficiency in high-concurrency scenarios.

[0050] Based on any of the above embodiments, in Embodiment 2 of this application, before step S10, steps A11 to A13 are further included: Step A11: Based on the resource data captured by the monitoring probe according to a preset period, resource indicator data is generated by encapsulating the business tag vector corresponding to the resource data.

[0051] In this embodiment, a monitoring probe refers to a data acquisition component deployed on a computing node, used to collect information on system resource usage. A preset period refers to a pre-configured time interval used to periodically trigger data collection. Resource data refers to the raw usage indicators of hardware resources such as CPU, memory, and network obtained through the monitoring probe. A business tag vector refers to a feature vector that encodes business-related attributes into numerical form. Encapsulation refers to combining and formatting data from different sources into a unified data structure. Resource indicator data refers to the final standardized data packet containing resource usage information and business tags.

[0052] As an optional implementation, the monitoring probe captures resource data from the operating system kernel interface at a preset period, while simultaneously querying business tag information through the application programming interface and converting the captured resource data into one-hot encoded vectors. Subsequently, the probe loads the resource data and business tag vectors into a memory buffer, performs serialization processing using a protocol buffer encoding library, adds timestamps and sequence number metadata, and generates a resource indicator data message packet. This method provides efficient serialization and deserialization performance, reducing network bandwidth consumption.

[0053] As an alternative implementation, the monitoring probe reads resource data at fixed time intervals and retrieves a predefined business tag mapping table from the local cache, converting the tags into numerical vectors. The probe then directly concatenates the resource data and vectors into a byte array, adding header information and a checksum using a custom binary format to form a resource indicator data block. This method is suitable for resource-constrained environments or offline data collection, offering simple processing but limited real-time performance and scalability.

[0054] Step A12: Preprocess the resource indicator data within each preset period according to the preset processing rules to generate the target sequence stream.

[0055] In this embodiment, the preset processing rules refer to a set of predefined data cleaning and transformation steps. Preprocessing refers to the cleaning, transformation, and normalization operations performed on the raw data. The target sequence stream refers to the standardized time series data generated after preprocessing that meets the model input requirements.

[0056] As an optional implementation, resource indicator data is retrieved from the message queue at a preset period, loaded into a memory buffer, and its integrity is checked and the location of missing values ​​is marked. Subsequently, an interpolation algorithm is used to fill in the missing values ​​with a neighbor-weighted approach, while a filtering algorithm based on the moving average of the sliding window is employed to suppress abnormal spikes. The processed data is then sent to a normalization module, where each dimension of the indicators is linearly mapped to a unified numerical range based on pre-calculated historical extreme values. The standardized data is then organized chronologically into standardized time series data. This method offers high flexibility, allowing for easy adjustment of preprocessing rules and parameter settings through software updates to adapt to different data characteristics and processing needs.

[0057] As an alternative implementation, resource indicator data is acquired through a direct memory access interface. A missing value processing unit identifies data gaps and fills them using a hardware interpolation algorithm. An anomaly suppression unit calculates statistics within a sliding window in real time and smoothly replaces data exceeding a threshold. Subsequently, a normalization unit performs linear transformation calculations through a built-in extreme value register, and finally outputs the processed data as standardized time series data through a high-speed serial interface. This method boasts extremely high processing speed and energy efficiency. All operations are performed on dedicated hardware circuits, eliminating the need for operating system scheduling and software instruction execution, thus achieving extremely low processing latency.

[0058] Step A13: Extract the features of the target sequence stream, and concatenate the obtained latent vector with the sine and cosine periodic codes to generate a conditional feature vector.

[0059] In this embodiment, the latent vector refers to the high-dimensional abstract feature representation extracted through a neural network. Sine and cosine periodic encoding refers to the periodic feature vector obtained by converting timestamp information using sine and cosine functions.

[0060] As an optional implementation, the target sequence stream is loaded into the input buffer of the accelerator card. The convolutional computation unit uses a multi-scale one-dimensional convolutional kernel group to perform sliding window computation on the input target sequence stream, and outputs a feature map after processing by a nonlinear activation function. Simultaneously, a periodic encoder calculates the corresponding sine and cosine periodic components based on the data packet timestamp information and combines them into an encoded vector. This feature map is then fed into a pooling layer for dimensionality reduction to obtain a latent vector. Finally, this latent vector and the periodic encoded vector are joined in a concatenation circuit to generate a conditional feature vector. This method offers extremely low latency and high energy efficiency, making it suitable for online service scenarios with stringent response time requirements.

[0061] As an alternative implementation, after the target sequence stream is transferred from host memory to GPU memory, a convolutional kernel function is invoked to perform multi-level feature extraction on the sequence. The output of each convolutional layer undergoes batch normalization and activation function processing, and a latent vector is generated through global pooling. Simultaneously, another computational unit computes sine and cosine periodic codes in parallel based on timestamp metadata. Finally, a matrix concatenation function is used to concatenate the latent vector and the periodic code vector along the feature dimension, and the resulting conditional feature vector is returned to host memory. This method is flexible and scalable, suitable for R&D experimental environments where feature extraction strategies require frequent optimization.

[0062] For example, in a scenario where a generative adversarial network (GAN) predicts computing power demand, a monitoring probe retrieves CPU, GPU, and network resource usage data from server nodes at fixed time intervals. Simultaneously, it obtains business information such as job type and SLA level from the cluster management system and encodes it into one-hot vectors. Subsequently, the probe encapsulates the resource data and business tag vectors into resource indicator data in Protocol Buffer format and transmits it to the data processing layer via a message middleware. In the preprocessing module, the resource indicator data undergoes missing value imputation, outlier suppression, and normalization to generate a standardized target sequence stream. This target sequence stream is then input to a feature extraction module, which extracts multi-scale features using a temporal convolutional network and gated recurrent units, outputting a latent vector. Simultaneously, timestamp information is mapped using sine and cosine functions to generate a periodic encoded vector. Finally, the latent vector and the periodic encoded vector are concatenated to form a conditional feature vector, which serves as the conditional input to the generator in the GAN, used to generate a predicted computing power demand sequence for future time periods.

[0063] By using a generative adversarial network framework, the problem of traditional prediction models failing to capture load mutations and periodic changes is solved, thus improving prediction accuracy and resource scheduling efficiency.

[0064] Based on any of the above embodiments, in Embodiment 3 of this application, referring to Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the control method for the computing power demand prediction system of this application. Step A13 further includes steps B11 to B13: Step B11: Divide the target sequence stream into segments using a sliding window of a preset fixed length to generate continuous time series segments.

[0065] In this embodiment, the preset fixed length refers to a pre-set, immutable window size parameter. Sliding window segmentation refers to the operation of moving along the sequence and truncating subsequences at fixed step sizes. A continuous time series segment refers to a fixed-length data subsequence block truncated in chronological order.

[0066] As an optional implementation, the target sequence stream is written to a circular buffer in the order of data arrival. When the amount of data accumulated in the buffer reaches a preset fixed length, the window controller triggers a segment truncation operation, copying a continuous fixed-length data from the current starting position in the buffer to the output buffer to form a time series segment. Subsequently, the window controller moves the starting position by a preset step size, discarding the oldest data and receiving new data. When the new data fills the space required by the step size, truncation is triggered again, and so on to generate continuous time series segments. This method has high memory efficiency, reusing memory space through the circular buffer, and is suitable for continuous data stream processing.

[0067] As an alternative implementation, a fixed-size memory block is pre-allocated for each time series segment, and an index mapping table is established. At the start of processing, the algorithm directly calculates the start and end positions of the window based on the current timestamp and a preset fixed length. It then directly accesses the corresponding data segment in the target sequence stream using memory address calculation, copies it to the pre-allocated memory block to form the time series segment. Subsequently, the index value is updated at fixed steps, the next window position is calculated, and the same operation is performed to generate all segments in batches, obtaining the time series segments. This method is computationally efficient, requires no maintenance of complex buffer states, and is suitable for batch processing of existing complete datasets.

[0068] Step B12: Extract sequence features of the time series segment through convolution calculation and gated loop operation, and output the latent vector representing the sequence features.

[0069] In this embodiment, convolution computation refers to the operation of sliding a convolution kernel across the input data to perform linear computation and feature extraction. Gated recurrent computation refers to recurrent neural network computation that controls the flow of information through a gating mechanism. Latent vectors refer to high-dimensional, abstract feature representations extracted through neural networks.

[0070] As an optional implementation, time-series segments are loaded from memory into an input buffer. The convolutional computation core uses multiple sets of one-dimensional convolutional kernels to perform sliding window computation on the sequence, extracting local feature patterns through linear multiplication-accumulation operations and nonlinear activation functions, and outputting a feature mapping tensor. This tensor is fed into a gate-controlled loop computation core, which controls the retention and forgetting of historical information through update and reset gates, performing temporal modeling of the sequence and outputting a final state vector. The final state vector is then passed through a fully connected layer for dimensionality transformation and compression, generating a fixed-dimensional latent vector. This method is computationally efficient and consumes little power, making it suitable for online inference scenarios with extremely high real-time requirements.

[0071] As an alternative implementation, feature extraction is achieved through parallel computation. The time-series segment is loaded into GPU memory. First, a parallel convolution kernel function is invoked, using multiple thread blocks to simultaneously compute convolution operations at different positions, generating a feature mapping matrix. This matrix is ​​then input into a gated recurrent unit kernel function, where parallelized matrix-vector operations and gated computation iteratively update the hidden state along the time steps. Finally, the hidden state at the last time step is extracted, and through another kernel function, it undergoes linear transformation and dimensionality reduction to output the latent vector. This method is flexible, highly programmable, and suitable for research and development environments where feature extraction algorithms require frequent adjustments and optimizations.

[0072] Step B13: Based on the timestamp information corresponding to the target sequence stream, calculate the daily and weekly periodic components through sine and cosine function mapping and concatenate them to generate the sine and cosine periodic code.

[0073] In this embodiment, sine and cosine function mapping refers to the calculation process of periodically transforming time values ​​using sine and cosine mathematical functions. The daily cycle component refers to the numerical characteristic representation reflecting the 24-hour cycle. The weekly cycle component refers to the numerical characteristic representation reflecting the 7-day cycle. Sine and cosine periodic encoding refers to the final generated composite vector representation containing cycle characteristic information.

[0074] As an optional implementation, timestamp information is extracted from the data packet header, and the clock parser converts it into two standardized time values: the number of seconds in a day and the number of seconds in a week. Subsequently, the trigonometric function calculation core calculates the corresponding sine and cosine function values ​​for these two time values ​​in parallel. The four sine and four cosine values ​​obtained are fed into a vector concatenation circuit, which combines them in a preset order to form an eight-dimensional periodic feature vector, thus creating a sine-cosine periodic code. This method exhibits extremely low computational latency and high energy efficiency, making it suitable for applications with strict latency requirements.

[0075] As an alternative implementation, timestamp data is read and converted to a standard time format. Sine and cosine functions from a mathematical library are then used to calculate the daily and weekly cycle components, respectively. Angle transformation is then applied to generate eight calculation results, which are stored in contiguous memory to form a cycle encoding vector. This method is simple to implement, highly flexible, and suitable for research and development environments where the cycle pattern needs frequent adjustments.

[0076] Step B14: Concatenate the sine and cosine periodic codes with the latent vector according to a preset vector concatenation method to generate the conditional feature vector.

[0077] In this embodiment, the preset vector concatenation method refers to the pre-defined operation rules for connecting and combining multiple vectors in the dimensional direction.

[0078] As an optional implementation, the sine and cosine periodic codes and the latent vectors are loaded into two independent regions of the input buffer. The concatenation controller reads a preset concatenation rule and then controls the data transport circuit to copy the data elements of the two vectors sequentially into the output register, forming a higher-dimensional conditional feature vector. This method provides high-speed feature fusion capabilities for real-time systems and is suitable for inference scenarios with strict processing speed requirements.

[0079] As an alternative implementation, the sine and cosine periodic codes and the latent vectors are stored in contiguous memory space. By calculating the target address offset, memory copy instructions are used to copy the data elements of the two vectors sequentially to the newly allocated memory block, generating a conditional feature vector. This method is simple to implement, highly flexible, and suitable for experimental environments where frequent adjustments to feature combination methods are required during the R&D and debugging phase.

[0080] For example, in a typical large-scale cloud-edge collaborative data center scenario, the data center consists of 5,000 physical host machines (each with dual Intel Xeon Silver 4316 processors, 2.3GHz clock speed, and 256GB of memory) and 320 NVIDIA A100 GPUs; the training cluster is located in an adjacent cold aisle server room, equipped with 4 × A100 80GB GPUs and HDR-200 gigabytes per second InfiniBand, serving as a point for offline training and online fine-tuning; the edge inference containers use 1 × NVIDIA T4 (or equivalent XPU / NPU) per node, running in Kubernetes 1.30, and elastically scaling using Docker images. The entire monitoring link is uniformly implemented through a 100GbE Spine-Leaf architecture, with a laboratory-tested one-way latency of <25µs. The software stack consists of Ubuntu 22.04 LTS, CUDA 11.8, cuDNN 8.9, and PyTorch 2.2. The acquisition-side probes use eBPF + OpenTelemetry1s for granular reporting, and the message bus uses Apache Pulsar 3.2. Data acquisition and message pipeline: Each host machine loads a lightweight BPF program in kernel mode to sample M=18 monitoring dimensions, including CPU utilization, GPU computation / memory usage, memory usage, and network throughput, every second. Let the original observation value of the m-th dimension at time t be... Simultaneously, job labels (JobType, SLA_Class, etc.) are extracted from the scheduling plane and encoded into one-hot vectors. The probe will and After being framed and encapsulated into a Protocol-Buffer, the data is pushed to the Pulsar topic metrics-raw. The average message size is 640 bytes, with a peak write rate of 8 gigabytes per second, easily covering peak scenarios. Preprocessing and feature construction: The receiving Kafka-Connect Sink writes metrics-raw to the distributed column-oriented storage ClickHouse and triggers real-time streaming preprocessing. For any monitoring sequence... Missing interpolation: If If there is a vacancy, the exponentially weighted mean of the most recent k=5 valid samples is used: ; Items with insufficient beginning or end are filled by mirroring. Anomaly suppression: Calculate the median within a sliding window of w=600 (10 min). and deviation .like Treated as a spike, replaced with Normalization, performing linear scaling on each dimension. The range is limited to [-1, 1], where Based on 60 days of full statistics, Use 10 to the power of negative cube to prevent division by zero. Then use a sliding window sequence of length w = 300s. Input convolution gated recursive encoder: The encoder consists of three layers of 1-D Dilated Temporal Convolution (convolutions {3,5,7}, dilation rates 1→2→4) stacked with one layer, outputting a 128-dimensional latent vector. To explicitly inject periodic information, the timestamp t is mapped to a D=16-dimensional vector using sine and cosine mapping. The splicing results in the following characteristics: .

[0081] By introducing multi-scale feature extraction and periodic coding mechanisms, the problem of traditional prediction models' inability to capture short-term fluctuations and long-term periodic patterns is solved, thus improving prediction accuracy and robustness.

[0082] Based on any of the above embodiments, in Embodiment 4 of this application, step S20 includes steps C11 to C13: Step C11: Based on the dependencies of the modeling features, perform residual connections, layer normalization, and feedforward network transformation on the high-dimensional latent space representation in sequence to output the primary hierarchical feature representation.

[0083] In this embodiment, the dependency relationship of modeling features refers to the interrelationship and influence patterns between feature elements. Residual connection refers to a skip connection operation that adds the input and output of a layer in a neural network. Layer normalization refers to standardizing the mean and variance of the output of a single layer in a neural network. Feedforward network transformation refers to nonlinear feature transformation performed through fully connected layers and activation functions. Primary layer feature representation refers to the basic feature vector after preliminary processing.

[0084] As an optional implementation, the high-dimensional latent space representation vector is read and stored in an input register. This vector is then simultaneously fed into the current processing layer and the residual connection buffer. The current processing layer performs a linear transformation on the vector, and its output is element-wise added to the original input in the buffer in an adder to achieve residual connections. The addition result is fed into the layer normalization core, where the mean and variance of the vector are calculated and standardized. The standardized vector is then fed into the feedforward network transformation module, where it undergoes two linear transformation layers and activation function processing to output the primary-level feature representation. This method is fast, energy efficient, and suitable for production environments with strict real-time requirements.

[0085] Step C12 involves iteratively refining the initial feature representation through feature processing, progressively refining the features and capturing multi-scale time dependencies, and outputting an intermediate-level feature representation.

[0086] In this embodiment, iterative feature processing refers to the process of gradually optimizing features through repeated computational steps. Primary feature representation refers to feature vectors that have undergone preliminary processing but are not yet fully refined. Progressive feature refinement refers to the operation of continuously improving feature quality and discriminative power through successive processing steps. Capturing multi-scale temporal dependencies refers to extracting temporal correlation patterns within different time ranges. Intermediate-level feature representation refers to more representative and abstract feature vectors obtained after multiple rounds of feature refinement.

[0087] As an optional implementation, the primary feature representation is loaded into the input register, followed by an iterative processing loop: In the first processing stage, the feature vector is processed by a temporal convolutional unit to extract local temporal patterns, and the output is fed into a short-term dependency modeling unit. In the second processing stage, the output of the previous stage is processed by a gated recurrent unit to capture intermediate temporal dependencies, and important time steps are highlighted through attention weight calculation. In the third processing stage, dilated convolutional operations are used to expand the receptive field to capture long-term dependency patterns, and the output of each stage is fused with the features of the previous stage through residual connections. After multiple rounds of iterative processing, the final output is a mid-level feature representation. This method has high processing efficiency and fully optimized temporal characteristics, making it suitable for scenarios with strict requirements for feature quality.

[0088] Step C13: Map the intermediate feature representation to the target dimension based on the linear projection layer, and decode and generate the resource demand prediction sequence according to the numerical constraint range.

[0089] In this embodiment, the linear projection layer refers to a neural network layer that performs linear transformations using a weight matrix and a bias vector. The intermediate feature representation refers to the abstract feature vector obtained after multi-level feature extraction and refinement. The numerical constraint range refers to the range of values ​​within which the output value needs to be restricted. Decoding refers to the process of converting the abstract feature representation into a specific predicted value.

[0090] As an optional implementation, the intermediate feature representation is loaded into the input register, and the pre-trained projection weight matrix and bias vector are read from the weight memory. A vector multiplier performs multiplication of the input vector with the weight matrix and adds the bias vector to the result, completing the linear projection transformation. The output vector is then fed into an activation function processor, which selects an appropriate activation function based on a preset numerical constraint range for nonlinear transformation, ensuring that the output value falls within the target interval, thus generating a resource demand prediction sequence. This method provides extreme performance for real-time prediction systems and is suitable for applications with stringent response speed requirements.

[0091] As an alternative implementation, the intermediate feature representation is loaded into the input buffer of the processing core. A linear algebra library function is then called to perform matrix multiplication, multiplying the input vector by the projection weight matrix stored in memory and adding a bias term. Subsequently, a numerical constraint processing function is called to scale and offset the result vector according to a preset output range, ensuring it strictly falls within a specified numerical interval, thus obtaining the resource demand prediction sequence. This method is suitable for development and experimental environments that require frequent adjustments to output characteristics and constraints.

[0092] For example, in a typical large-scale cloud-edge collaborative data center scenario, the adversarial prediction network of this system adopts the classic generative adversarial network (GAN) framework, including a generator. Discriminator Generator Designed as a six-layer Transformer-Decoder structure, each layer includes a multi-head self-attention mechanism and a feedforward neural network module to capture the temporal dependencies and feature interaction patterns of sequence data. The input includes a standard normally distributed random noise vector z~N(0, I_64) and real-time observed multidimensional conditional features. The two are first fused into a unified representation through an affine transformation, and then decoded layer by layer through a Transformer-Decoder network to gradually generate a multidimensional load prediction sequence for the next T=60 seconds: .

[0093] By employing multi-level feature refinement and multi-scale dependency capture mechanisms, the problem of insufficient representation of complex temporal patterns in traditional prediction models is solved, thereby improving prediction accuracy and temporal consistency.

[0094] Based on any of the above embodiments, in Embodiment 5 of this application, step S30 includes steps D11 to D13: Step D11: Based on the resource demand prediction sequence and the actual load sequence, calculate the residual between the predicted value and the actual value point by point according to the preset quantiles, and generate the quantile residual sequence.

[0095] In this embodiment, the actual load sequence refers to the sequence of resource usage values ​​actually collected by the monitoring system. Preset quantiles refer to pre-defined probability quantile values, including 50%, 90%, and 99%. Point-by-point calculation refers to the operation of calculating each data element at a corresponding position in the sequence individually. The residual between the predicted and actual values ​​refers to the difference between the predicted and actual values. The quantile residual sequence refers to the sequence formed by arranging the residual values ​​calculated according to specific quantiles in chronological order.

[0096] As an optional implementation, the resource demand prediction sequence and the actual load sequence are loaded into the input buffer, and each quantile processing channel independently reads the preset quantile parameters. Then, each channel performs residual calculation in parallel for each time point in the sequence. When the predicted value is greater than the actual value, the residual at that time point is calculated as the quantile parameter multiplied by the difference. When the predicted value is less than the actual value, the residual is calculated as the quantile parameter minus one, multiplied by the difference. The residual calculation results for all time points are organized chronologically to form a quantile residual sequence. This method provides high quantile residual calculation capabilities for real-time prediction systems and is suitable for online learning scenarios with strict requirements on computational latency.

[0097] Step D12: Determine and suppress abnormal residual values ​​in the quantile residual sequence by using a preset threshold, and output a robust residual metric.

[0098] In this embodiment, the preset threshold refers to a pre-defined numerical limit used to determine whether a data point is abnormal. Suppression refers to the processing operation of restricting or replacing abnormal values. Abnormal residual values ​​refer to residual data points that significantly deviate from the normal range. Robust residual metric refers to a statistically robust residual evaluation index obtained after outlier processing.

[0099] As an optional implementation, the quantile residual sequence is loaded into the input buffer, and a sliding window statistician calculates the median and interquartile range of the residual sequence within the current window in real time. Then, according to preset anomaly detection rules, each residual value in the sequence is compared to a threshold; when a value exceeds the threshold range, the outlier replacement module replaces it with the median value of the current window. The processed sequence is then sent to the metric calculation unit, where a robust residual metric is generated using a weighted average algorithm. This method provides stable outlier suppression capabilities for real-time systems, ensuring the reliability of the residual metric.

[0100] As an alternative implementation, the system loads the quantile residual sequence into the processor's data memory. First, a sorting function is called to sort the current data block and calculate statistics. Then, a dynamic threshold is determined based on a preset outlier probability. The processor iterates through each data point in the sequence, replacing outliers exceeding the dynamic threshold with a local weighted average. Finally, the root mean square value of the processed sequence is calculated as the output robust residual metric. This method is suitable for R&D environments where outlier handling strategies require frequent optimization and can quickly verify the effectiveness of different algorithms.

[0101] Step D13: The robust residual metric is weighted and fused according to the quantile weight coefficient to generate the composite prediction error value.

[0102] As an optional implementation, the quantile weight coefficient vector is read, and a robust residual metric vector is loaded from the data buffer. The multiplier performs parallel multiplication of the residual metric value at each quantile with its corresponding weight coefficient, generating a weighted result vector. The accumulator sums all weighted results and outputs the composite prediction error value. This method provides real-time prediction systems with high-speed and stable error calculation capabilities, ensuring a rapid response from the prediction model.

[0103] As an alternative implementation, the quantile weights and robust residual metrics are loaded into different regions of the vector register. The processor calls the vector multiplication instruction to perform element-wise multiplication, generating a temporary product vector. Then, the vector accumulation instruction is used to sum all elements in the temporary vector to obtain the composite prediction error value. This method is suitable for R&D and experimental environments, enabling rapid testing of the impact of different weight settings on model performance.

[0104] For example, in a typical large-scale cloud-edge collaborative data center scenario, the discriminator A deep discriminant module based on a five-layer Temporal Convolutional Network (TCN) is employed, with spectral normalization applied to each layer to ensure training stability and prevent mode collapse. The discriminator is input with a real payload sequence. Or generate a load sequence After successive temporal convolutions and nonlinear activations, a scalar score is output to measure the authenticity of the input sequence. The network optimization process is based on Wasserstein distance and a gradient penalty term, forming the Wasserstein-GP loss function, which is defined as follows: ; in, Represents real samples With generated samples Random linear interpolation points between, penalty coefficient The value is set to 10. To further improve the prediction accuracy of the generated sequence and reduce point prediction errors, the generator of this system also introduces the Huber-Quantile composite loss function: ; in, This represents the Huber loss function. quantile coefficient Adjusting the weighting coefficients The aforementioned losses are intended to improve the model's ability to predict extreme load changes and capture the risk of abnormal fluctuations, thereby ensuring the robustness of the predictions.

[0105] By introducing quantile residual calculation and outlier suppression mechanisms, the problem of traditional prediction models being sensitive to extreme values ​​and noise is solved, thus improving the robustness and accuracy of predictions.

[0106] Based on any of the above embodiments, in Embodiment Six of this application, step S40 includes steps E11 to E13: Step E11: The resource demand prediction sequence after data normalization and the actual load sequence are recombined into a time series tensor format to generate a standardized input tensor.

[0107] In this embodiment, data normalization refers to the standardization process of mapping the original data to a specific numerical range through linear transformation. Reorganizing into a time-series tensor format means reorganizing the sequence data into a multi-dimensional array structure according to time step, feature dimension, and other axes. Standardized input tensors refer to tensor data that meets the model input requirements after normalization and structuring.

[0108] As an optional implementation, the normalized prediction sequence and the actual load sequence are loaded into the input buffer. A data format converter rearranges the data elements of the two sequences according to preset time-series dimension parameters. Data is written into a specific memory layout pattern according to tensor dimension requirements, where the time step dimension uses contiguous storage and the feature dimension uses interleaved storage, generating a standardized input tensor. This method offers high data access efficiency and optimized memory layout, making it suitable for production environments with stringent throughput requirements.

[0109] As an alternative implementation, the normalized sequence data is stored in a contiguous memory space. The tensor processing core defines the shape and stride parameters of the tensor by configuring dimension descriptors. A memory-mapped view is created based on the descriptors, and a standardized input tensor is generated by virtual dimension recombination without actually moving the data. This method is suitable for R&D and experimental environments that require frequent adjustments to the input format and can quickly adapt to different model structure requirements.

[0110] Step E12: By calculating the causal convolution of the input tensor and superimposing the spectral normalization constraint weight matrix on the causal convolution, a multidimensional high-level feature tensor is generated.

[0111] In this embodiment, the input tensor refers to a structured multidimensional data array after preprocessing and standardization. Causal convolution is a special convolution operation that ensures the output depends only on the current and previous inputs, preventing the leakage of future information. Spectral normalization constraint weight matrix refers to a regularization technique that enhances model stability by restricting the spectral norm of the weight matrix. Multidimensional high-level feature tensor refers to a tensor representation containing high-level semantic information obtained after feature extraction and transformation.

[0112] As an optional implementation, the input tensor is loaded into the accelerator's on-chip cache. The causal convolution computation unit employs a padding strategy to ensure the output length matches the input, and performs constrained convolution kernel sliding computation in the time dimension. The result of the convolution computation is fed into the spectral normalization module, which approximates the maximum singular value of the weight matrix using an iterative method. This value is then used to scale and constrain the weight matrix, outputting a multidimensional high-level feature tensor. This method provides high-speed and stable feature extraction capabilities for inference systems and is suitable for deployment scenarios with strict requirements on latency and power consumption.

[0113] Step E13: Based on the global average pooling operation on the multidimensional high-level feature tensor, the obtained long-range features are mapped to scalar scores through a fully connected layer to generate the scalar scores that characterize the authenticity of the sequence.

[0114] In this embodiment, global average pooling refers to the computational process of averaging all elements in each feature channel to compress the spatial dimension. A fully connected layer is a neural network layer in which each neuron is connected to all neurons in the previous layer.

[0115] As an optional implementation, a multidimensional high-level feature tensor is loaded into the input buffer. A global pooling calculator performs parallel averaging on all spatial elements of each feature channel, compressing the high-dimensional tensor into a long-range feature vector. This vector is then fed into a fully connected processing core, which reads the pre-trained weight matrix and bias vector, and performs matrix multiplication and bias addition. The calculated result is then passed through a linear activation function to generate the final scalar score. This method provides high-speed and stable score output for real-time discrimination systems, making it suitable for online inference scenarios with strict response speed requirements.

[0116] For example, in a scenario where a generative adversarial network predicts computational demand, the system first reconstructs the resource demand prediction sequence (after data normalization) and the actual load sequence into a time-series tensor format, generating a standardized input tensor. Then, by calculating a causal convolution operation on this input tensor and superimposing a spectral normalization constraint weight matrix on the convolution result, a multidimensional high-level feature tensor containing advanced semantic features is generated. Finally, based on a global average pooling operation on this feature tensor, the spatial dimension is compressed, and the obtained long-range features are mapped to scalar scores through a fully connected layer, generating a discriminator output score representing the authenticity of the sequence.

[0117] By introducing a feature extraction mechanism based on causal convolution and spectral normalization, the problems of gradient instability and limited receptive field in traditional discrimination models for long sequence processing are solved, thereby improving the accuracy and stability of sequence authenticity discrimination.

[0118] Based on any of the above embodiments, in Embodiment Seven of this application, step S50 includes steps F11 to F13: Step F11: The composite prediction error value and the scalar score are weighted and summed according to preset weight coefficients to generate a joint optimization loss value.

[0119] In this embodiment, the preset weighting coefficient refers to a pre-set proportional parameter used to adjust the importance of different loss items. The joint optimization loss value refers to the value of a unified objective function that integrates the composite prediction error value and the scalar score.

[0120] As an optional implementation, preset weight coefficients are read from the configuration memory and loaded into two areas of the weight register, while the composite prediction error value and scalar score are input into the data buffer. The multiplier performs parallel multiplication of the composite prediction error value with its corresponding weight, and simultaneously performs multiplication of the scalar score with its corresponding weight. The two products are fed into an accumulator for summation to generate the joint optimization loss value. This method provides the training system with high-speed and stable loss calculation capabilities, ensuring the efficiency and stability of large-scale training.

[0121] As an alternative implementation, preset weight coefficients are stored in a designated memory location, and the composite prediction error value and scalar score are loaded into different channels of a vector register. The processor executes vector multiplication instructions, multiplying each of the two input values ​​by its corresponding weight coefficient. Then, scalar addition instructions are used to sum the two products to obtain the joint optimization loss value. This method is suitable for research and experimental environments that require frequent adjustments to the loss weights, and can quickly explore the impact of different loss combinations on model performance.

[0122] Step F12: Based on the input of the joint optimization loss value, the parameter gradients of the generator and discriminator are calculated alternately at a preset update frequency.

[0123] In this embodiment, the preset update frequency refers to the pre-defined ratio of the number of times the generator and discriminator parameters are updated. Alternating computation refers to a computation mode in which parameters are executed in turn according to a specific order. The generator refers to the neural network model responsible for generating the predicted sequence. The discriminator refers to the neural network model responsible for evaluating the authenticity of the sequence. The parameter gradient refers to the direction and magnitude of the model parameter updates calculated through the backpropagation algorithm.

[0124] As an optional implementation, the joint optimization loss value is loaded into the loss register, and the update scheduler configures the current computation mode according to a preset update frequency. During the discriminator update cycle, the gradient engine initiates backpropagation computation on the discriminator parameters, automatically differentiating the partial derivatives of the loss value with respect to the weights of each discriminator layer to generate the discriminator parameter gradients and storing them in the gradient buffer. During the generator update cycle, the gradient engine switches to the generator parameter path, calculating the partial derivatives of the loss value with respect to the weights of each generator layer to generate the generator parameter gradients. This method provides stable and efficient dual-path gradient computation for large-scale training and is suitable for production environments with fixed update frequencies.

[0125] Step F13 involves processing the parameter gradients through gradient synchronization to drive model parameter updates and output the trained prediction generator model.

[0126] In this embodiment, gradient synchronization refers to the operation of coordinating and aggregating parameter gradients across multiple computing devices. Parameter gradients are the set of numerical values ​​representing the direction and magnitude of model weight updates, calculated through backpropagation. Driving model parameter updates refers to the process of adjusting neural network weights using gradient values ​​through an optimization algorithm.

[0127] As an optional implementation, the calculated parameter gradients are stored in a local gradient buffer. The communication controller establishes connections with corresponding units of other computing nodes via a high-speed interconnect network, and uses the All-Reduce algorithm to globally sum and average the gradients of all nodes. The synchronized global gradients are sent to the optimizer interface, which executes an optimization algorithm based on preset learning rate and momentum parameters, calculates the model parameter update amount, and writes the new weights into the model memory. When the training iterations reach a preset number or the loss value converges, the system exports the final driving model parameters from memory to persistent storage, updates the driving model parameters, and outputs the trained prediction generator model. This method provides excellent scalability and efficiency for ultra-large-scale training and is suitable for training ultra-large generator models.

[0128] As an alternative implementation, each compute node sends its locally computed parameter gradients to a central parameter server. The server aggregates all gradients and performs an averaging operation. The optimizer computes global parameter updates on the server side and then broadcasts the updated weights to all compute nodes. Nodes receive the new weights and update their local model copies, repeating the iteration until the model converges. Finally, stable driving model parameters are exported from the parameter server. The driving model parameters are updated, and the trained prediction generator model is output. This method is suitable for medium-sized clusters or heterogeneous computing environments, providing better visualization and control of the training process.

[0129] For example, combining the above loss terms, the overall optimization objective of the model can be expressed as a joint optimization problem: ; To achieve better training results, the training process is divided into three stages: Warm-up: In the initial stage, a teacher-mandated training strategy is used to gradually approach the true target until the mean absolute percentage error (MAPE) is less than 15%. Adversarial Alternating Optimization Stage: In this stage, the discriminator and generator are alternately optimized at a 5:1 update ratio, meaning the discriminator is updated five times for every generator update, to maintain training stability and prevent premature convergence or pattern collapse. Simultaneously, early stopping is performed every 10 epochs using the validation set to prevent overfitting. Online Fine-tuning Stage: A sliding data window mechanism is used to fine-tune network weights in real time based on the latest load sequence to adapt to dynamically changing data distribution and further improve prediction accuracy. The model training in this system adopts a distributed strategy, utilizing the All-Reduce operator from the NVIDIA NCCL library for gradient synchronization, with a single training batch size of 512. The optimizer uses the AdamW algorithm, with an initial learning rate set to... Furthermore, a cosine annealing scheduling strategy is employed to ensure smooth and effective convergence during training. Concept drift detection and adaptive calibration: Real-time calculation of residuals during the online inference phase. Standardize it as And input Page-Hinkley flow statistics: ; when Distribution drift was identified. To avoid impacting online service, only the first four layers of the generator decoder were frozen, while the last two layers and the output header were retained for a 32-step mini-batch hot update; all gradient constraints were applied. Internally designed to prevent violent oscillations. Actual measurements show that MAPE stabilizes within 60 seconds.

[0130] By jointly optimizing the loss function and using an alternating training mechanism, the problems of training instability and mode collapse in generative adversarial networks are solved, thereby improving the model's convergence speed and prediction accuracy.

[0131] Based on any of the above embodiments, in Embodiment 8 of this application, after step S50, steps G11~G13 are further included: Step G11: Based on the multidimensional load sequence input to the prediction generator model, the decoder is called to perform parallel computation, outputting the original prediction tensor for the future time period, and generating a preset quantile prediction curve through the quantile calculation kernel.

[0132] In this embodiment, the multidimensional load sequence refers to a data set containing resource usage data in multiple dimensions arranged chronologically. The raw prediction tensor refers to the multidimensional prediction result directly output by the neural network without post-processing. The quantile calculation kernel refers to a processing unit specifically used to calculate different probability quantiles. The preset quantile prediction curve refers to a curve formed by arranging the predicted values ​​generated according to specified probability quantiles in chronological order.

[0133] As an optional implementation, a multidimensional load sequence is loaded from host memory into the prediction generator model parameters. A parallel computing architecture is used to simultaneously execute the decoder's multi-layer attention mechanism and feedforward network computation to generate the original prediction tensor for future time periods. This tensor is then transmitted to a quantile calculation engine, which has multiple built-in quantile processing channels. Each channel computes the predicted value corresponding to a specified quantile in parallel, and an interpolation algorithm is used to generate a smooth quantile prediction curve. This method provides high-speed and stable quantile prediction capabilities for real-time prediction systems, making it suitable for online decision-making scenarios with stringent response speed requirements.

[0134] Step G12: Based on the prediction curve of the first preset quantile and the actual load sequence, the residual calculation kernel is called to generate a real-time error sequence, and the prediction generator model is calibrated based on the real-time error sequence.

[0135] In this embodiment, the first preset quantile refers to a pre-defined probability quantile. The real-time error sequence refers to the set of prediction error values ​​arranged in chronological order. Calibration refers to the optimization process of improving prediction accuracy by adjusting model parameters.

[0136] As an optional implementation, the predicted curve of the first preset quantile and the true load sequence are loaded into the input buffer. The residual calculation module calculates the absolute difference between the predicted value and the true value at each time point in parallel, generating a real-time error sequence and storing it in the error buffer. Subsequently, the model calibration engine reads this error sequence, calculates the error statistics through a sliding window, and when the average error exceeds a preset threshold, triggers the incremental learning algorithm. This freezes the pre-feature extraction layer of the prediction generator model, optimizes only the weight parameters of the output layer using gradient descent, updates the model parameters using the error sequence as training samples, and writes the updated driving model parameters back to the model memory to complete the calibration. This method provides a stable and reliable model calibration capability for high-precision prediction systems and is suitable for scenarios with strict requirements for prediction consistency.

[0137] As an alternative implementation, the predicted curve and the actual sequence are transmitted to the processing unit, where a residual calculation kernel function is invoked to generate an error sequence in parallel. The calibration controller monitors the statistical characteristics of the error sequence, and when a persistent deviation is detected, it initiates a model calibration procedure: selecting the error data from the most recent time window as the calibration dataset, employing an online learning algorithm to adjust the weights of the last layer of the generator model, and reducing prediction bias through iterative optimization. The calibrated model is then immediately used for the next round of prediction. This method is suitable for application scenarios with variable environments and can quickly adapt to changes in data distribution.

[0138] Step G13: Based on the calibrated prediction generator model, update the prediction curve of the second preset quantile to generate a flexible scaling decision.

[0139] In this embodiment, the calibrated prediction generator model refers to a sequence prediction neural network whose accuracy has been improved after error feedback and parameter optimization. The second preset quantile refers to a pre-set second probability quantile. Flexible scaling decision refers to a strategy instruction that automatically adjusts the scale of computing resources based on the prediction results.

[0140] As an optional implementation, a calibrated prediction generator model receives real-time input data and performs forward inference computation to output an original prediction sequence. The quantile calculation core reads the second preset quantile parameter, calculates the predicted value sequence for that quantile using sorting and interpolation algorithms, and generates an updated prediction curve. The strategy logic module applies predefined rules based on this curve to generate a flexible scaling decision instruction, which is sent to the resource scheduler via the system bus to execute resource adjustments. This method provides nanosecond-level response capabilities for high-frequency decision-making systems, ensuring the timeliness and consistency of resource adjustments.

[0141] As an alternative implementation, the calibrated model is deployed in the inference service, receiving prediction requests through a preset interface. The model service returns the raw prediction data. The quantile calculation service calls statistical library functions to calculate the prediction curve for the second preset quantile. The decision engine service parses the prediction curve according to the policy rules in the configuration file, generates scaling decisions, and sends them to the executor component via message middleware. This method is suitable for development and deployment environments where policies require frequent optimization and testing, and can support multi-tenancy and complex policy logic.

[0142] For example, refer to Figure 4 , Figure 4This is the overall flowchart of this application. The data acquisition module acquires a multidimensional load sequence, which is input to the adversarial prediction module. The module's built-in decoder performs parallel computation, outputting the raw prediction tensor for the future time period. A first preset quantile prediction curve is generated using the quantile calculation kernel. The dynamic drift detection and adaptive calibration module, based on this prediction curve and the actual load sequence, calls the residual calculation kernel to generate a real-time error sequence and calibrates the parameters of the adversarial prediction model based on this sequence. The decision output module updates the second preset quantile prediction curve using the calibrated model. The scheduling system generates elastic scaling decisions based on this curve, while the monitoring feedback module and energy consumption estimator form a closed-loop optimization.

[0143] By combining the adversarial prediction module and the dynamic drift detection and adaptive calibration module, the problem of insufficient accuracy of traditional prediction models under data distribution drift and complex load modes is solved, thereby improving prediction accuracy and resource scheduling efficiency.

[0144] This application provides a computing power demand prediction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the control method of the computing power demand prediction system in the above embodiment 1.

[0145] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing a computing power demand prediction device according to embodiments of this application. The computing power demand prediction device in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, scheduling execution devices, personal digital assistants (PDAs), tablet computers (PADs), portable multimedia players (PMPs), energy efficiency management devices, etc., as well as fixed terminals such as resource scheduling platforms, desktop computers, etc. Figure 5 The computing power demand prediction device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0146] like Figure 5As shown, the computing demand prediction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the computing demand prediction device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the computing demand forecasting device to communicate wirelessly or wiredly with other devices to exchange data. Although a computing demand forecasting device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0147] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0148] The computing power demand prediction device provided in this application, employing the control method of the computing power demand prediction system in the above embodiments, can solve the technical problem of high server idle rate caused by excessive reserved resources, leading to low resource utilization. Compared with the prior art, the beneficial effects of the computing power demand prediction device provided in this application are the same as the beneficial effects of the control method of the computing power demand prediction system provided in the above embodiments, and other technical features in this computing power demand prediction device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0149] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0150] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0151] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the control method of the computing power demand prediction system in the above embodiments.

[0152] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0153] The aforementioned computer-readable storage medium may be included in the computing demand forecasting device; or it may exist independently and not be assembled into the computing demand forecasting device.

[0154] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the computing power demand prediction device, the computing power demand prediction device performs the following actions: concatenates the input conditional features with random Gaussian noise to obtain a high-dimensional latent space representation; sequentially decodes the high-dimensional latent space representation, arranges it according to the decoding order, and determines a resource demand prediction sequence for future time periods; calculates a weighted loss based on the resource demand prediction sequence and the actual load sequence according to preset quantization points, and outputs a composite prediction error value; extracts long-range features from the resource demand prediction sequence and the actual load sequence, and obtains a scalar score through feature scalarization; and weightedly fuses the composite prediction error value and the scalar score to construct a joint optimization objective, thereby alternately updating the generator and discriminator, and iteratively generating a prediction generator model.

[0155] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0156] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0157] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0158] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the control method of the above-described computing power demand prediction system. This solves the technical problem of high server idle rates caused by excessive reserved resources, leading to low resource utilization. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the control method of the computing power demand prediction system provided in the above embodiments, and will not be elaborated upon here.

[0159] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A control method for a computing power demand prediction system, characterized in that, The control of the computing power demand prediction system includes: The conditional features of the input are concatenated with random Gaussian noise and fused to obtain a high-dimensional latent space representation; The high-dimensional latent space representation is decoded sequentially, and the resource demand prediction sequence for future time periods is determined according to the decoding order. Based on the resource demand prediction sequence and the actual load sequence, a weighted loss is calculated according to a preset quantile, and a composite prediction error value is output. Long-range features of the resource demand prediction sequence and the actual load sequence are extracted, and scalar scores are obtained through feature scalarization. The composite prediction error value and the scalar score are weighted and fused to construct a joint optimization objective, which alternately updates the generator and discriminator to iteratively generate the prediction generator model.

2. The control method for the computing power demand prediction system as described in claim 1, characterized in that, Before the step of fusing the conditional features of the concatenated input with random Gaussian noise to obtain a high-dimensional latent space representation, the control method of the computing power demand prediction system further includes: Based on the resource data captured by the monitoring probe according to a preset period, the resource indicator data is generated by encapsulating the business tag vector corresponding to the resource data. Preprocess the resource indicator data within each preset period according to preset processing rules to generate a target sequence stream; The features of the target sequence stream are extracted, and the obtained latent vector is concatenated with the sine and cosine periodic codes to generate a conditional feature vector.

3. The control method for the computing power demand prediction system as described in claim 2, characterized in that, The steps of extracting features from the target sequence stream and concatenating the obtained latent vector with sine and cosine periodic codes to generate conditional feature vectors include: The target sequence stream is segmented by a sliding window of a preset fixed length to generate continuous time series segments. The sequence features of the time series segment are extracted by convolution calculation and gated loop operation, and the latent vector representing the sequence features is output. Based on the timestamp information corresponding to the target sequence stream, the daily and weekly periodic components are calculated and concatenated by sine and cosine function mapping to generate the sine and cosine periodic code. The conditional feature vector is generated by concatenating the sine and cosine periodic codes with the latent vector according to a preset vector concatenation method.

4. The control method for the computing power demand prediction system as described in claim 1, characterized in that, The steps of decoding the high-dimensional latent space representation of the sequence and determining the resource demand prediction sequence for future time periods according to the decoding order include: Based on the dependencies of the modeling features, the high-dimensional latent space representation is sequentially subjected to residual connections, layer normalization, and feedforward network transformation to output the primary-level feature representation. The initial feature representation is processed iteratively to refine the features and capture multi-scale time dependencies, outputting an intermediate-level feature representation. The intermediate feature representation is mapped to the target dimension based on a linear projection layer, and the resource demand prediction sequence is generated by decoding according to the numerical constraint range.

5. The control method for the computing power demand prediction system as described in claim 1, characterized in that, The step of calculating a weighted loss based on the resource demand prediction sequence and the actual load sequence according to a preset quantile, and outputting a composite prediction error value, includes: Based on the resource demand prediction sequence and the actual load sequence, the residual between the predicted value and the actual value is calculated point by point according to the preset quantile to generate the quantile residual sequence. By identifying and suppressing abnormal residual values ​​in the quantile residual sequence through a preset threshold, a robust residual metric is output. The composite prediction error value is generated by weighting and fusing the robust residual metric according to the quantile weight coefficient.

6. The control method for the computing power demand prediction system as described in claim 1, characterized in that, The step of extracting long-range features from the resource demand forecast sequence and the actual load sequence, and obtaining a scalar score through feature scalarization processing includes: The resource demand prediction sequence after data normalization and the actual load sequence are recombined into a time series tensor format to generate a standardized input tensor. By calculating the causal convolution of the input tensor and superimposing a spectral normalization constraint weight matrix on the causal convolution, a multidimensional high-level feature tensor is generated. Based on the global average pooling operation on the multidimensional high-level feature tensor, and by mapping the obtained long-range features to scalar scores through a fully connected layer, the scalar scores representing the authenticity of the sequence are generated.

7. The control method for the computing power demand prediction system as described in claim 1, characterized in that, The steps of weightedly fusing the composite prediction error value and the scalar score to construct a joint optimization objective, and iteratively generating the prediction generator model by alternately updating the generator and discriminator, include: The composite prediction error value and the scalar score are weighted and summed according to preset weighting coefficients to generate a joint optimization loss value. Based on the input of the joint optimization loss value, the parameter gradients of the generator and discriminator are calculated alternately at a preset update frequency; The parameter gradients are processed by gradient synchronization to drive model parameter updates and output the trained prediction generator model.

8. The control method for the computing power demand prediction system as described in claim 1, characterized in that, After the step of weightedly fusing the composite prediction error value and the scalar score to construct a joint optimization objective, and iteratively generating the prediction generator model by alternately updating the generator and discriminator, the control method of the computing power demand prediction system further includes: The prediction generator model is based on a multidimensional load sequence input, which is called to perform parallel computation, outputting the original prediction tensor for the future time period, and generating a preset quantile prediction curve through the quantile calculation kernel. Based on the predicted curve of the first preset quantile and the actual load sequence, the residual calculation kernel is called to generate a real-time error sequence, and the prediction generator model is calibrated based on the real-time error sequence. Based on the calibrated prediction generator model, the prediction curve of the second preset quantile is updated to generate flexible scaling decisions.

9. A computing power demand prediction device, characterized in that, The computing power demand prediction device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the control method for the computing power demand prediction system as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the control method of the computing power demand prediction system as described in any one of claims 1 to 8.

Images