A method and apparatus for predicting the load of intelligent computing centers based on task element feature modulation
By using an adaptive wavelet packet decomposition and dynamic weight modulation state-space model, the problems of task-driven mechanism and multiple time scales in load forecasting of intelligent computing centers are solved, achieving high-precision and stable load forecasting, which is suitable for power load forecasting in intelligent computing centers.
Patent Information
- Application Number
- CN202611000876.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-07
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for load forecasting in intelligent computing centers lack characterization of task-driven mechanisms, making it difficult to uniformly handle different forecast time scales. Furthermore, they lack methods for dynamically adjusting the contributions of multi-source features and compensating for load structure migration errors, leading to inaccurate forecasts and decreased stability.
An adaptive wavelet packet decomposition method is used to decompose the total load sequence of the intelligent computing center into multiple levels. The dynamic weights are determined by the normalized mutual information between the sub-band component sequences and the task element features. A state-space model is constructed for baseline prediction, and corrections are made through low-frequency gain calibration and high-frequency offset compensation mechanisms. Finally, the time-domain load prediction results are reconstructed.
It significantly improves prediction accuracy and robustness in mixed load scenarios, can adaptively match load changes at different time scales, provides task-level attribution analysis, and has good engineering applicability and promotion value.
Smart Images

Figure CN123115631A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system load forecasting technology, specifically to a method and apparatus for intelligent computing center load forecasting based on task element feature modulation. Background Technology In recent years, the rapid growth of large-scale model training and online inference services has driven a continuous increase in the electricity consumption of intelligent computing centers, while also giving their power loads distinct new characteristics compared to traditional data centers. On the one hand, large-scale training operations generate high-power, long-duration loads on the cluster side, which may be accompanied by significant power and thermal fluctuations, placing higher constraints on power supply and cooling systems. Relevant public research and engineering practices have pointed out that the power fluctuations and frequency components of training loads pose new challenges to the stability of data center infrastructure and the power grid.
[0002] On the other hand, inference operations are driven by user requests, exhibiting sudden and random characteristics. Load is more prone to short-term spikes and rapid fluctuations, demonstrating strong volatility on a smaller time scale. For this type of AI data center load, existing research has emphasized that its power consumption curve may fluctuate rapidly, and load patterns are significantly affected by changes in the underlying algorithms and computing tasks, making predictions relying solely on historical statistical patterns difficult to reliably adapt.
[0003] With the rapid growth of computing demands in artificial intelligence (AI), data centers, especially Artificial Intelligence Data Centers (AIDCs) designed for AI training and inference tasks, have become a significant component of new power loads. AIDCs commonly adopt an integrated training and inference deployment model, running model training and inference services simultaneously within the same physical cluster. While this hybrid load model improves GPU resource utilization, it presents a significant challenge to power load forecasting. Extensive research has been conducted in academia and engineering on the problem of data center power demand forecasting. Generally, existing technologies can be categorized into three types: single-scale forecasting methods based on historical load sequences, load forecasting methods based on time-frequency decomposition or spectrum enhancement, and complex time-series forecasting methods based on multimodal fusion or attention mechanisms.
[0004] The first category of methods is single-scale forecasting based on historical power load sequences. These methods typically use historical power data as the core input, combining it with time labels, environmental variables, or some operational indicators to construct a feature set. Then, they utilize statistical learning models or deep learning models to predict future load. Typical methods include Autoregressive Moving Average (ARIMA) models, Seasonal Autoregressive Moving Average (SARIMA) models, and deep learning models widely used in recent years, such as Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), Temporal Convolutional Networks (TCNs), and Transformers. These methods predict future load changes by learning the time dependencies and periodic patterns in historical load sequences, and have been widely applied in traditional industrial load, commercial building load, and conventional data center load forecasting. In engineering practice, such methods typically follow a relatively fixed process: first, historical load curves are collected and a time series dataset is constructed; then, time features (such as hours, days of the week, and seasons) and a small number of operational indicators (such as temperature and equipment utilization) are extracted; then, the above-mentioned statistical models or deep learning models are trained; finally, load forecast results under a fixed forecast step size or a fixed forecast window are output, and the model is optimized and evaluated using indicators such as mean squared error (MSE), mean absolute error (MAE), or mean absolute percentage error (MAPE).
[0005] The second category of methods is load forecasting based on time-frequency decomposition or spectrum enhancement. Addressing the common non-stationarity and multi-scale characteristics of power load sequences, some studies have proposed performing frequency domain decomposition or time-frequency analysis on the load sequence before modeling to enhance the model's ability to represent complex signal structures. Typical methods include Fast Fourier Transform (FFT), Wavelet Transform, Empirical Mode Decomposition (EMD), and Variational Mode Decomposition (VMD). These methods decompose the original load sequence into several sub-sequences with different frequency characteristics, such as trend components, periodic components, and high-frequency disturbance components, which are then modeled separately or jointly, thereby improving prediction accuracy. Compared to methods that rely solely on time-domain modeling, time-frequency decomposition methods can more clearly capture the periodic structure and multi-scale fluctuation characteristics of power loads, offering advantages when dealing with load sequences with significant periodicity or multi-frequency characteristics. Therefore, this type of method is widely used in power system load forecasting, energy demand forecasting, and some data center power forecasting research.
[0006] The third category of methods comprises complex time series forecasting methods based on multimodal fusion or attention mechanisms. With the continuous improvement of data center operation monitoring systems, more and more auxiliary information can be used for load forecasting, such as equipment operation logs, resource utilization, task scheduling records, and environmental sensor data. Therefore, some studies attempt to jointly model time series data with multi-source information, improving forecasting capabilities through multimodal fusion frameworks. Common methods include utilizing attention mechanisms or gating structures to achieve information interaction between different features, such as multi-head attention mechanisms, gated fusion networks, and cross-modal alignment structures. These methods, by introducing additional information sources, alleviate the problem of insufficient information from a single time series to some extent and can improve the model's expressive power in complex operating environments. Therefore, they have been applied in operation and maintenance monitoring, industrial process forecasting, and some data center operation analysis scenarios.
[0007] Although the above three methods have made some progress in data center load forecasting research, they still have significant limitations under the mixed load conditions of training and inference in intelligent computing centers.
[0008] First, existing methods generally rely on historical load extrapolation as their core modeling approach, lacking a characterization of task-driven mechanisms. Compared to traditional data centers, intelligent computing centers exhibit significantly different load formation mechanisms. Artificial intelligence training tasks typically represent sustained loads with long operating times and high computational intensity, while inference services exhibit short-term, high-frequency, pulse-like loads. Load changes are not only influenced by historical power sequences but are also closely related to factors such as task type, model size, parallelization strategies, and changes in service traffic. When training tasks switch, model sizes are adjusted, or inference service traffic surges, the load structure may change significantly. In such cases, simply relying on historical load sequences for extrapolation prediction fails to reflect the new load formation mechanisms in a timely manner, leading to inaccurate characterization of sustained base load, delayed response to short-term peak loads, and decreased prediction stability.
[0009] Secondly, most existing methods model for a single prediction time scale, lacking a unified multi-time scale prediction mechanism. In most studies, prediction models are typically trained for fixed time steps (e.g., minutes or hours) and output prediction results at the corresponding time scale. However, in the operational scenario of intelligent computing centers, different types of tasks have significantly different impacts on the time scale. For example, training tasks focus more on long-term operational trends and continuous baseload, while inference operations focus more on short-term fluctuations and sudden load changes. If a single time scale modeling approach or a simple multi-model combination method is used, it is often difficult to simultaneously capture short-term fluctuations and predict medium- to long-term trends within the same framework, thus generating systematic errors at different prediction scales.
[0010] Furthermore, existing multimodal forecasting methods mostly employ static strategies for feature fusion, lacking dynamic adjustment mechanisms tailored to the forecasting objective. While some studies attempt to incorporate multi-source information into load forecasting models, this information is typically fused through simple concatenation or fixed weighting. The importance of different types of features varies significantly across different forecasting time scales; for example, short-term forecasting relies more on local time-domain fluctuations, while long-term forecasting depends more on trend structures or cyclical features. Static fusion methods struggle to dynamically adjust the contributions of various features according to the forecasting objective, thus limiting the model's adaptability in multi-time-scale forecasting scenarios. Moreover, in situations with frequent task switching and rapid load structure evolution, existing methods typically lack a dedicated mapping relationship between "task-driven information and power load changes," and also lack mechanisms to compensate for structural migration errors.
[0011] In summary, while existing technologies provide an important foundation for historical load modeling, time-frequency feature extraction, and multi-source information fusion, they still have the following shortcomings: First, there is a lack of specialized modeling methods for the power load characteristics of intelligent computing centers under mixed training and inference load conditions; second, there is a lack of a prediction mechanism that can uniformly handle different prediction time scales; and third, there is a lack of effective methods for dynamically adjusting the contributions of multi-source features and compensating for load structure migration errors based on the prediction target. Therefore, it is necessary to propose a new load forecasting method that can integrate time-domain features, frequency-domain features, and task-side driving information within a unified framework, thereby achieving accurate multi-time-scale power load forecasting for intelligent computing centers. Summary of the Invention
[0012] To overcome the above-mentioned shortcomings, this invention proposes a method and apparatus for predicting the load of intelligent computing centers based on task element feature modulation.
[0013] Firstly, a method for predicting the load of intelligent computing centers based on task element feature modulation is provided, the method comprising: An adaptive wavelet packet decomposition method is used to perform multi-level decomposition of the total load sequence of the intelligent computing center, resulting in multiple sub-band component sequences. The dynamic weights of the subband component sequence in the training domain and the dynamic weights of the inference domain are determined based on the normalized mutual information between the subband component sequence and the features of each task element. A state-space model for benchmark prediction of subbands within the prediction interval is constructed based on the subband component sequences and their corresponding dynamic weights in the training and inference domains. The state-space model is used to perform baseline predictions for each sub-band within the prediction interval, resulting in a baseline prediction sequence for each sub-band within the prediction interval. Based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence, the baseline prediction sequence of each sub-band in the prediction interval is corrected to obtain the corrected baseline prediction sequence of each sub-band in the prediction interval. Perform inverse wavelet transform on the benchmark prediction sequence of each sub-band within the prediction interval after correction, and reconstruct the benchmark prediction sequence of each sub-band within the prediction interval into the time-domain load prediction result.
[0014] Preferably, during the process of using the adaptive wavelet packet decomposition method to perform multi-level decomposition of the total load sequence of the intelligent computing center, when the current decomposition depth k satisfies When k is used as the maximum decomposition depth, where, For the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of the nth frequency band node and the th Layer wavelet packet decomposition layer 1 The sum of the logarithmic energy entropy of each frequency band node For the first Layer wavelet packet decomposition layer 1 Logarithmic energy entropy of each frequency band node.
[0015] Furthermore, the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of each frequency band node is as follows:
[0016] In the above formula, For the first Layer wavelet packet decomposition layer 1 Number of discrete data points per frequency band node For the first Layer wavelet packet decomposition layer 1 The i-th discrete data point of a frequency band node It is a positive smoothing factor.
[0017] Preferably, the task meta-features include at least one of the following: resource allocation features, training task features, inference task features, interconnection communication features, and scheduling event features.
[0018] Furthermore, the resource allocation features include the number of GPUs allocated, the number of CPU cores, the amount of video memory used, and the amount of RAM used; the training task features include model parameter size, training status identifier, gradient synchronization event identifier, distributed parallel configuration, and checkpoint saving event; the inference task features include request arrival rate, request queue length, input / output sequence length, model loading status, and changes in online service instances; and the interconnection and communication features include inter-node network throughput, inter-chip bandwidth utilization, and communication blocking flags.
[0019] Preferably, the normalized mutual information between the sub-band component sequences and the features of each task element is as follows:
[0020] In the above formula, Subband component sequence With task meta-features Normalized mutual information between them Subband component sequence With task meta-features Mutual information between them Subband component sequence Information entropy Task meta-features Information entropy.
[0021] Furthermore, the dynamic weights of the sub-band component sequence in the training domain and the dynamic weights in the inference domain are as follows:
[0022]
[0023] In the above formula, The training domain dynamic weights for the sub-band component sequences. The training domain association score is used for the sub-band component sequence. , The inference domain association score for the sub-band component sequence. , For the dynamic weights of the inference domain of the sub-band component sequences, The set of training domain features in the task meta-features. The first feature is used to aggregate weights. This is the set of inference domain features in the task meta-features. The aggregate weights are for the second feature.
[0024] Furthermore, the state-space model is as follows:
[0025]
[0026] In the above formula, For sub-band i at time i The hidden state vector. For sub-band i at time i Discrete state transition parameters, For sub-band i at time i Discrete input mapping parameters, For sub-band i at time i -1 hidden state vector For sub-band i at time i sequence of components, For sub-band i at time i The baseline forecast value, For sub-band i at time i The time-varying mapping parameters, Input the pass-through term for sub-band i.
[0027] Furthermore, the sub-band i at time... The discrete state transition parameters and discrete input mapping parameters are as follows:
[0028] In the above formula, For sub-band i at time i The basic state transition matrix, For sub-band i at time i Time-varying input mapping parameters in continuous state space, For sub-band i at time i The discretization step size parameter, This is a discretization processing function.
[0029] Furthermore, the sub-band i at time... The time-varying input mapping parameters and subband i in the continuous state space at time... The time-varying mapping parameters are as follows:
[0030] In the above formula, For a moment Gated generation function, Generate functions for input mapping. For the output mapping generation function, where the time... The gating generation function is as follows:
[0031] The input mapping generation function and the output mapping generation function are as follows: ,
[0032] In the above formula, It is a non-linear activation function. For gating weights, For gated bias. It is a comprehensive feature vector composed of the current input, the local context features of the current input, and dynamically assigned weights. Let be the first projection matrix. For the first bias term, This is the second projection matrix. This is the second bias term.
[0033] Furthermore, the sub-band i at time... The discretization step size parameters are as follows:
[0034] In the above formula, For activation function, This is the weight matrix. For the third bias term, This is the first positive adjustment coefficient corresponding to subband i. This is the second positive adjustment coefficient corresponding to subband i. For sub-band i at time i The dynamic weights of the training domain For sub-band i at time i Dynamic weights of the inference domain.
[0035] Furthermore, before correcting the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval, the process includes: If subband i is at time i If the difference between the dynamic weights of the inference domain and the dynamic weights of the training domain is greater than or equal to 1, then subband i is classified as a high-frequency transient subband; otherwise, subband i is classified as a low-frequency steady-state subband.
[0036] Furthermore, for the low-frequency steady-state subband, the subband m at time t is determined by the following formula. Gain compensation coefficient:
[0037] In the above formula, For sub-band m at time Gain compensation coefficient, As the lower limit of the calibration coefficient, To ensure a safe upper limit for the calibration coefficient, Let m be the linear mapping weights for subband m. For subband m at time steady-state scale-type feature input, The linear mapping bias for subband m For sub-band m at time Compensation training domain dynamic weights, For sub-band m at time The dynamic weights of the compensation inference domain; Subband m at time Baseline forecast Revised to .
[0038] Furthermore, for high-frequency transient subbands, the subband m at time t is determined by the following formula. Offset compensation term:
[0039] In the above formula, For m at time Offset compensation term, , , These are the learnable linear mapping weights corresponding to traffic surge features, concurrent pressure features, and communication pulse features, respectively. For sub-band m at time Corresponding traffic surge characteristics For sub-band m at time The corresponding concurrent pressure characteristics, For sub-band m at time Corresponding communication pulse characteristics, This is the third bias term; Subband m at time Baseline forecast Revised to .
[0040] Furthermore, the sub-band m at time The process of obtaining the dynamic weights of the compensated training domain and the dynamic weights of the compensated inference domain includes: The dynamic weights of the subband component sequence compensation training domain and the dynamic weights of the compensation inference domain are determined based on the normalized mutual information between the subband component sequence and the meta-features of each compensation task.
[0041] Furthermore, the compensation task meta-features include: traffic surge features, concurrent pressure features, communication pulse features, and resource switching features.
[0042] Furthermore, the traffic surge characteristics include: active parameter scale, number of GPUs allocated, video memory usage, and number of training instances; the concurrency pressure characteristics include: request arrival rate, change in request arrival rate, queue length, and sudden increase in queue length; the communication pulse characteristics include: gradient synchronization flag, model loading event, pipeline switching event, and sudden increase in communication bandwidth; and the resource switching characteristics include: load change event type.
[0043] Secondly, a task-based feature modulation-based intelligent computing center load prediction device is provided, the task-based feature modulation-based intelligent computing center load prediction device comprising: The decomposition module is used to perform multi-level decomposition of the total load sequence of the intelligent computing center using the adaptive wavelet packet decomposition method to obtain multiple sub-band component sequences. The determination module is used to determine the dynamic weights of the training domain and the dynamic weights of the inference domain of the sub-band component sequence based on the normalized mutual information between the sub-band component sequence and the features of each task element. The building module is used to construct a state-space model for benchmark prediction of subbands within the prediction interval based on the subband component sequence and its corresponding dynamic weights in the training and inference domains. The prediction module is used to perform baseline predictions for each sub-band within the prediction interval using the state space model, and obtain the baseline prediction sequence for each sub-band within the prediction interval. The correction module is used to correct the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence, so as to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval. The analysis module is used to perform inverse wavelet transform on the benchmark prediction sequence of each sub-band within the prediction interval, and reconstruct the benchmark prediction sequence of each sub-band within the prediction interval into a time-domain load prediction result.
[0044] Thirdly, a computer device is provided, comprising: one or more processors; The processor is used to store one or more programs; When the one or more programs are executed by the one or more processors, the intelligent computing center load prediction method based on task meta-feature modulation is implemented.
[0045] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed, the aforementioned intelligent computing center load prediction method based on task element feature modulation is implemented.
[0046] The above-described technical solutions of the present invention have at least one or more of the following beneficial effects: This invention provides a method and apparatus for predicting the load of intelligent computing centers based on task-based feature modulation, which has the following advantages compared to existing technologies: First, by using adaptive wavelet packet decomposition and mutual information dynamic weight generation, the aliasing features of training and inference tasks in the time-frequency domain are effectively decoupled, significantly improving the prediction accuracy in mixed load scenarios. Second, the selective state-space model based on dynamically allocated weight modulation can adaptively match different time scales of low-frequency steady-state and high-frequency transient sub-bands, enhancing the model's robustness to extreme conditions such as task switching and traffic bursts. Third, the low-frequency gain calibration and high-frequency offset compensation mechanism endow the prediction results with clear physical interpretability, providing task-level attribution analysis for scheduling decisions. Finally, experimental verification shows that this invention achieves superior accuracy and stability compared to existing mainstream methods at different prediction step sizes, demonstrating good engineering applicability and promotion value. Attached Figure Description
[0047] Figure 1 This is a schematic diagram of the main steps of the intelligent computing center load prediction method based on task element feature modulation according to an embodiment of the present invention; Figure 2 This is a diagram showing the actual load sequence of the intelligent computing center during the test period of this invention and its typical sub-band results after adaptive wavelet packet decomposition. Figure 3 This is a normalized mutual information heatmap between typical low-frequency / high-frequency subbands and different task element feature categories in an embodiment of the present invention. Detailed Implementation
[0048] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0050] Example 1 See appendix Figure 1 , Figure 1 This is a schematic flowchart illustrating the main steps of a smart computing center load prediction method based on task element feature modulation, according to an embodiment of the present invention. Figure 1 As shown, the intelligent computing center load prediction method based on task element feature modulation in this embodiment of the invention mainly includes the following steps: Step S101: The total load sequence of the intelligent computing center is decomposed into multiple sub-band component sequences by using the adaptive wavelet packet decomposition method. Step S102: Determine the dynamic weights of the subband component sequence in the training domain and the dynamic weights of the inference domain based on the normalized mutual information between the subband component sequence and the features of each task element; Step S103: Construct a state-space model for benchmark prediction of subbands within the prediction interval based on the subband component sequence and its corresponding dynamic weights in the training and inference domains; Step S104: Use the state space model to perform baseline prediction for each sub-band within the prediction interval to obtain the baseline prediction sequence for each sub-band within the prediction interval. Step S105: Based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence, the baseline prediction sequence of each sub-band in the prediction interval is corrected to obtain the corrected baseline prediction sequence of each sub-band in the prediction interval. Step S106: Perform inverse wavelet transform on the modified sub-band's reference prediction sequence within the prediction interval to reconstruct the time-domain load prediction result.
[0051] Conventional wavelet packet decomposition typically employs a fixed decomposition depth or a simple pruning strategy based on subband energy proportions. This faces two major limitations in intelligent computing center scenarios: First, when the load structure undergoes a transition (such as the completion of a training task, the launch of a new model, or a sudden surge in inference traffic), the fixed depth cannot adapt to the new time-frequency characteristics. Second, the energy threshold is not sensitive enough to sparse but large-amplitude transient pulses, which can easily disperse the pulse energy into multiple subbands, weakening the accuracy of subsequent task attribution analysis.
[0052] To this end, the present invention introduces logarithmic energy entropy as a criterion for node splitting, and adopts the entropy reduction splitting criterion to achieve dynamic adaptation of the decomposition path.
[0053] In this embodiment, during the process of using the adaptive wavelet packet decomposition method to perform multi-level decomposition of the total load sequence of the intelligent computing center, when the current decomposition depth k satisfies When k is used as the maximum decomposition depth, where, For the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of the nth frequency band node and the th Layer wavelet packet decomposition layer 1 The sum of the logarithmic energy entropy of each frequency band node For the first Layer wavelet packet decomposition layer 1 Logarithmic energy entropy of each frequency band node.
[0054] In one implementation, the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of each frequency band node is as follows:
[0055] In the above formula, For the first Layer wavelet packet decomposition layer 1 Number of discrete data points per frequency band node For the first Layer wavelet packet decomposition layer 1 The i-th discrete data point of a frequency band node It is a positive smoothing factor.
[0056] In this embodiment, the task meta-features include at least one of the following: resource allocation features, training task features, inference task features, interconnection communication features, and scheduling event features.
[0057] In one implementation, the resource allocation characteristics include the number of GPUs allocated, the number of CPU cores, the amount of video memory used, and the amount of RAM used; the training task characteristics include the model parameter scale, training status identifier, gradient synchronization event identifier, distributed parallel configuration, and checkpoint saving event; the inference task characteristics include the request arrival rate, request queue length, input / output sequence length, model loading status, and changes in online service instances; and the interconnection communication characteristics include inter-node network throughput, inter-chip bandwidth utilization, and communication blocking flags.
[0058] In this embodiment, the normalized mutual information between the sub-band component sequence and each task element feature is as follows:
[0059] In the above formula, Subband component sequence With task meta-features Normalized mutual information between them Subband component sequence With task meta-features Mutual information between them Subband component sequence Information entropy Task meta-features Information entropy.
[0060] In one implementation, the sub-band component sequence training domain dynamic weights and inference domain dynamic weights are as follows:
[0061]
[0062] In the above formula, The training domain dynamic weights for the sub-band component sequences. The training domain association score is used for the sub-band component sequence. , The inference domain association score for the sub-band component sequence. , For the dynamic weights of the inference domain of the sub-band component sequences, The set of training domain features in the task meta-features. The first feature is used to aggregate weights. This is the set of inference domain features in the task meta-features. The aggregate weights are for the second feature.
[0063] In one implementation, the state-space model is as follows:
[0064]
[0065] In the above formula, For sub-band i at time i The hidden state vector. For sub-band i at time i Discrete state transition parameters, For sub-band i at time i Discrete input mapping parameters, For sub-band i at time i -1 hidden state vector For sub-band i at time i sequence of components, For sub-band i at time i The baseline forecast value, For sub-band i at time i The time-varying mapping parameters, Input the pass-through term for sub-band i.
[0066] Unlike traditional fixed-parameter state-space models, in this step, at least one or more of the discretization step size, input mapping strength, and output response weights in the state update process can be modulated in real time based on the current sub-band input, local context features, and dynamically allocated weights. This allows the model to retain stronger long-term memory capabilities in low-frequency steady-state sub-bands and enhance its sensitivity to sudden disturbances in high-frequency transient sub-bands.
[0067] In one implementation, the sub-band i at time The discrete state transition parameters and discrete input mapping parameters are as follows:
[0068] In the above formula, For sub-band i at time i The basic state transition matrix, For sub-band i at time i Time-varying input mapping parameters in continuous state space, For sub-band i at time i The discretization step size parameter, This is a discretization processing function.
[0069] In one implementation, the sub-band i at time The time-varying input mapping parameters and subband i in the continuous state space at time... The time-varying mapping parameters are as follows:
[0070] In the above formula, For a moment Gated generation function, Generate functions for input mapping. For the output mapping generation function, where the time... The gating generation function is as follows:
[0071] The input mapping generation function and the output mapping generation function are as follows: ,
[0072] In the above formula, It is a non-linear activation function. For gating weights, For gated bias. It is a comprehensive feature vector composed of the current input, the local context features of the current input, and dynamically assigned weights. Let be the first projection matrix. For the first bias term, This is the second projection matrix. This is the second bias term.
[0073] In one implementation, the sub-band i at time The discretization step size parameters are as follows:
[0074] In the above formula, For activation function, This is the weight matrix. For the third bias term, This is the first positive adjustment coefficient corresponding to subband i. This is the second positive adjustment coefficient corresponding to subband i. For sub-band i at time i The dynamic weights of the training domain For sub-band i at time i Dynamic weights of the inference domain.
[0075] In one implementation, before correcting the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval, the process includes: If subband i is at time i If the difference between the dynamic weights of the inference domain and the dynamic weights of the training domain is greater than or equal to 1, then subband i is classified as a high-frequency transient subband; otherwise, subband i is classified as a low-frequency steady-state subband.
[0076] In one implementation, for the low-frequency steady-state subband, the subband m at time t is determined by the following formula. Gain compensation coefficient:
[0077] In the above formula, For sub-band m at time Gain compensation coefficient, As the lower limit of the calibration coefficient, To ensure a safe upper limit for the calibration coefficient, Let m be the linear mapping weights for subband m. For sub-band m at time steady-state scale-type feature input, The linear mapping bias for subband m For sub-band m at time Compensation training domain dynamic weights, For sub-band m at time The dynamic weights of the compensation inference domain; Subband m at time Baseline forecast Revised to .
[0078] In one implementation, for a high-frequency transient subband, the subband m at time t is determined by the following formula: Offset compensation term:
[0079] In the above formula, For m at time Offset compensation term, , , These are the learnable linear mapping weights corresponding to traffic surge features, concurrent pressure features, and communication pulse features, respectively. For sub-band m at time Corresponding traffic surge characteristics For sub-band m at time The corresponding concurrent pressure characteristics, For sub-band m at time Corresponding communication pulse characteristics, This is the third bias term; Subband m at time Baseline forecast Revised to .
[0080] In one implementation, the sub-band m at time The process of obtaining the dynamic weights of the compensated training domain and the dynamic weights of the compensated inference domain includes: The dynamic weights of the subband component sequence compensation training domain and the dynamic weights of the compensation inference domain are determined based on the normalized mutual information between the subband component sequence and the meta-features of each compensation task.
[0081] In one implementation, the compensation task meta-features include: traffic surge features, concurrent pressure features, communication pulse features, and resource switching features.
[0082] In one implementation, the traffic surge characteristics include: active parameter scale, number of GPUs allocated, video memory usage, and number of training instances; the concurrency pressure characteristics include: request arrival rate, change in request arrival rate, queue length, and sudden increase in queue length; the communication pulse characteristics include: gradient synchronization flag, model loading event, pipeline switching event, and sudden increase in communication bandwidth; and the resource switching characteristics include: load change event type.
[0083] In one specific implementation, since the above steps perform sub-band baseline prediction and compensation correction separately, under certain extreme conditions, the superposition results of individual sub-bands may lead to local prediction values exceeding reasonable physical ranges. Therefore, after completing the inverse transform reconstruction, physical constraint correction is performed on the time-domain prediction results to improve the usability and engineering rationality of the output results.
[0084] The physical constraint correction includes at least one or more of the following: 1) Non-negativity constraint. When the predicted load value is less than zero, it is truncated to zero or a preset minimum power value; 2) Capacity limit constraint. When the predicted load value exceeds the rated power supply capacity or safe operation limit of the target power supply unit, it will be cut off to the corresponding limit value; 3) Rate of change constraint. When the change in the predicted value between adjacent time points exceeds the maximum allowable load ramp rate of the system, a smoothing constraint is applied to this change; 4) Abnormal spike suppression. Non-physical spikes caused by the superposition of local high-frequency compensation can be corrected by sliding smoothing, amplitude limiting functions, or neighborhood consistency constraints.
[0085] In one specific implementation, this invention uses measured load data from a computing center for verification, spanning from March 1, 2024 to March 31, 2025, with a time resolution of 15 minutes. To ensure that the training, verification, and testing processes conform to the actual application patterns in time series prediction scenarios, the data is divided into training, verification, and testing sets in chronological order, with a ratio of 70%:20%:10%. The training set is used for model parameter learning, the verification set for model structure and hyperparameter tuning, and the testing set for final performance evaluation. The load sequence represents the total active load of the computing center. Task metadata is derived from the publicly available cluster-trace-v2026-spot-gpu dataset, and combined with monitoring logs, resource scheduling information, and business request records from the computing center, a task meta-feature set matching this invention is constructed. These task meta-features mainly include training scale features, training communication features, inference traffic features, inference queuing features, and resource switching features, used to characterize the driving effect of different task activities on load changes.
[0086] During the experiment, the original load sequence was first processed by imputing missing values, smoothing outliers, and normalizing. The task element feature sequences were then synchronized and aligned according to a unified time reference. Further, an adaptive wavelet packet decomposition method was used to perform multi-scale frequency band decomposition on the total load sequence. The decomposition depth and pruning path were automatically determined using the logarithmic energy entropy criterion, resulting in several low-frequency and high-frequency sub-bands. Subsequently, the sliding window normalized mutual information between each sub-band and the task element features was calculated, and this information was used to construct dynamic weight allocations for different frequency bands, thus providing a basis for subsequent sub-band prediction and compensation.
[0087] Figure 2The actual load sequence of the intelligent computing center during the test period and its typical sub-band results after adaptive wavelet packet decomposition are presented. As shown in the figure, the original total load sequence contains both a clear slowly varying baseline and local high-frequency disturbances. The low-frequency sub-band aa maintains a high degree of consistency with the original load in terms of overall trend, effectively representing the steady-state baseline formed by continuous training task operation and long-term resource occupation. The low-frequency sub-band ad retains some mid-to-low-frequency fluctuation information while maintaining a smooth trend. In contrast, the high-frequency sub-bands dd and da have smaller amplitudes but more drastic fluctuations, exhibiting obvious spikes, suddenness, and randomness, reflecting short-term load disturbances caused by factors such as sudden inference requests, queue jitter, communication pulses, and resource switching. These results demonstrate that the adaptive wavelet packet decomposition method used in this invention can effectively separate the low-frequency steady-state components from the high-frequency transient components in the mixed load of the intelligent computing center, thus providing a clear frequency band basis for subsequent differentiated prediction and compensation correction based on task element features.
[0088] Figure 3 Normalized mutual information heatmaps between typical low-frequency / high-frequency subbands and different task feature categories are presented. The heatmaps show that low-frequency subbands `low_band_1` and `low_band_2` have higher mutual information values with training scale and training communication features, but relatively weaker correlations with inference traffic and inference queuing features. Conversely, high-frequency subbands `high_band_1` and `high_band_2` show stronger correlations with inference traffic, inference queuing, and resource switching features. This indicates that low-frequency subbands are mainly driven by long-term training task execution and continuous resource consumption, while high-frequency subbands are mainly driven by bursts of inference traffic, changes in concurrent queuing, and resource switching events. These results further validate the rationality of the technical approach of "frequency band decomposition—task feature association—dynamic weight generation" in this invention. That is, different frequency bands not only differ significantly in their temporal morphology but also in their task-driving sources, thus providing a basis for subsequent selective state-space modeling and joint high- and low-frequency compensation.
[0089] As shown in Table 1, the proposed method exhibits superior prediction accuracy and stability under different prediction step sizes. For example, with prediction step sizes of 24, 48, 96, and 168, the mean squared error (MSE) of the proposed method is 0.122552, 0.151212, 0.197554, and 0.240346, respectively, and the mean absolute error (MAE) is 0.178170, 0.219678, 0.275411, and 0.316914, respectively. All of these are superior to the compared methods such as PatchTST, iTransformer, Autoformer, Informer, and DLinear. In short-term prediction scenarios (e.g., prediction step size of 24), the proposed method exhibits lower errors compared to PatchTST, iTransformer, and DLinear. As the prediction step size increases, the advantages of the proposed method are maintained, especially in medium- to long-term prediction scenarios (e.g., prediction step sizes of 96 and 168), where the error growth is relatively smaller, indicating that the present invention has better robustness in long-term time-series dependency modeling and load structure change adaptation. Meanwhile, Informer and Autoformer show significantly larger errors at various prediction step sizes, indicating that traditional methods based on unified time-series modeling struggle to simultaneously characterize steady-state trends and transient disturbances in the mixed load scenarios of intelligent computing centers. In summary, the error summary table verifies that the present invention, through the technical route of "band decoupling—selective state-space modeling—high- and low-frequency joint compensation," can effectively reduce prediction errors and maintain good generalization performance and stability at different prediction scales.
[0090] Table 1
[0091] Example 2 Based on the same inventive concept, the present invention also provides a smart computing center load prediction device based on task element feature modulation, the smart computing center load prediction device based on task element feature modulation includes: The decomposition module is used to perform multi-level decomposition of the total load sequence of the intelligent computing center using the adaptive wavelet packet decomposition method to obtain multiple sub-band component sequences. The determination module is used to determine the dynamic weights of the training domain and the dynamic weights of the inference domain of the sub-band component sequence based on the normalized mutual information between the sub-band component sequence and the features of each task element. The building module is used to construct a state-space model for benchmark prediction of subbands within the prediction interval based on the subband component sequence and its corresponding dynamic weights in the training and inference domains. The prediction module is used to perform baseline predictions for each sub-band within the prediction interval using the state space model, and obtain the baseline prediction sequence for each sub-band within the prediction interval. The correction module is used to correct the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence, so as to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval. The analysis module is used to perform inverse wavelet transform on the benchmark prediction sequence of each sub-band within the prediction interval, and reconstruct the benchmark prediction sequence of each sub-band within the prediction interval into a time-domain load prediction result.
[0092] Preferably, during the process of using the adaptive wavelet packet decomposition method to perform multi-level decomposition of the total load sequence of the intelligent computing center, when the current decomposition depth k satisfies When k is used as the maximum decomposition depth, where, For the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of the nth frequency band node and the th Layer wavelet packet decomposition layer 1 The sum of the logarithmic energy entropy of each frequency band node For the first Layer wavelet packet decomposition layer 1 Logarithmic energy entropy of each frequency band node.
[0093] Furthermore, the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of each frequency band node is as follows:
[0094] In the above formula, For the first Layer wavelet packet decomposition layer 1 Number of discrete data points per frequency band node For the first Layer wavelet packet decomposition layer 1 The i-th discrete data point of a frequency band node It is a positive smoothing factor.
[0095] Preferably, the task meta-features include at least one of the following: resource allocation features, training task features, inference task features, interconnection communication features, and scheduling event features.
[0096] Furthermore, the resource allocation features include the number of GPUs allocated, the number of CPU cores, the amount of video memory used, and the amount of RAM used; the training task features include model parameter size, training status identifier, gradient synchronization event identifier, distributed parallel configuration, and checkpoint saving event; the inference task features include request arrival rate, request queue length, input / output sequence length, model loading status, and changes in online service instances; and the interconnection and communication features include inter-node network throughput, inter-chip bandwidth utilization, and communication blocking flags.
[0097] Preferably, the normalized mutual information between the sub-band component sequences and the features of each task element is as follows:
[0098] In the above formula, Subband component sequence With task meta-features Normalized mutual information between them Subband component sequence With task meta-features Mutual information between them Subband component sequence Information entropy Task meta-features Information entropy.
[0099] Furthermore, the dynamic weights of the sub-band component sequence in the training domain and the dynamic weights in the inference domain are as follows:
[0100]
[0101] In the above formula, The training domain dynamic weights for the sub-band component sequences. The training domain association score is used for the sub-band component sequence. , The inference domain association score for the sub-band component sequence. , For the dynamic weights of the inference domain of the sub-band component sequences, The set of training domain features in the task meta-features. The first feature is used to aggregate weights. This is the set of inference domain features in the task meta-features. The aggregate weights are for the second feature.
[0102] Furthermore, the state-space model is as follows:
[0103]
[0104] In the above formula, For sub-band i at time i The hidden state vector. For sub-band i at time i Discrete state transition parameters, For sub-band i at time i Discrete input mapping parameters, For sub-band i at time i -1 hidden state vector For sub-band i at time i sequence of components, For sub-band i at time i The baseline forecast value, For sub-band i at time i The time-varying mapping parameters, Input the pass-through term for sub-band i.
[0105] Furthermore, the sub-band i at time... The discrete state transition parameters and discrete input mapping parameters are as follows:
[0106] In the above formula, For sub-band i at time i The basic state transition matrix, For sub-band i at time i Time-varying input mapping parameters in continuous state space, For sub-band i at time i The discretization step size parameter, This is a discretization processing function.
[0107] Furthermore, the sub-band i at time... The time-varying input mapping parameters and subband i in the continuous state space at time... The time-varying mapping parameters are as follows:
[0108] In the above formula, For a moment Gated generation function, Generate functions for input mapping. For the output mapping generation function, where the time... The gating generation function is as follows:
[0109] The input mapping generation function and the output mapping generation function are as follows: ,
[0110] In the above formula, It is a non-linear activation function. For gating weights, For gated bias. It is a comprehensive feature vector composed of the current input, the local context features of the current input, and dynamically assigned weights. Let be the first projection matrix. For the first bias term, This is the second projection matrix. This is the second bias term.
[0111] Furthermore, the sub-band i at time... The discretization step size parameters are as follows:
[0112] In the above formula, For activation function, This is the weight matrix. For the third bias term, This is the first positive adjustment coefficient corresponding to subband i. This is the second positive adjustment coefficient corresponding to subband i. For sub-band i at time i The dynamic weights of the training domain For sub-band i at time i Dynamic weights of the inference domain.
[0113] Furthermore, before correcting the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval, the process includes: If subband i is at time i If the difference between the dynamic weights of the inference domain and the dynamic weights of the training domain is greater than or equal to 1, then subband i is classified as a high-frequency transient subband; otherwise, subband i is classified as a low-frequency steady-state subband.
[0114] Furthermore, for the low-frequency steady-state subband, the subband m at time t is determined by the following formula. Gain compensation coefficient:
[0115] In the above formula, For sub-band m at time Gain compensation coefficient, As the lower limit of the calibration coefficient, To ensure a safe upper limit for the calibration coefficient, Let m be the linear mapping weights for subband m. For sub-band m at time steady-state scale-type feature input, The linear mapping bias for subband m For sub-band m at time Compensation training domain dynamic weights, For sub-band m at time The dynamic weights of the compensation inference domain; Subband m at time Baseline forecast Revised to .
[0116] Furthermore, for high-frequency transient subbands, the subband m at time t is determined by the following formula. Offset compensation term:
[0117] In the above formula, For m at time Offset compensation term, , , These are the learnable linear mapping weights corresponding to traffic surge features, concurrent pressure features, and communication pulse features, respectively. For sub-band m at time Corresponding traffic surge characteristics For sub-band m at time The corresponding concurrent pressure characteristics, For sub-band m at time Corresponding communication pulse characteristics, This is the third bias term; Subband m at time Baseline forecast Revised to .
[0118] Furthermore, the sub-band m at time The process of obtaining the dynamic weights of the compensated training domain and the dynamic weights of the compensated inference domain includes: The dynamic weights of the subband component sequence compensation training domain and the dynamic weights of the compensation inference domain are determined based on the normalized mutual information between the subband component sequence and the meta-features of each compensation task.
[0119] Furthermore, the compensation task meta-features include: traffic surge features, concurrent pressure features, communication pulse features, and resource switching features.
[0120] Furthermore, the traffic surge characteristics include: active parameter scale, number of GPUs allocated, video memory usage, and number of training instances; the concurrency pressure characteristics include: request arrival rate, change in request arrival rate, queue length, and sudden increase in queue length; the communication pulse characteristics include: gradient synchronization flag, model loading event, pipeline switching event, and sudden increase in communication bandwidth; and the resource switching characteristics include: load change event type.
[0121] Example 3 Based on the same inventive concept, this invention also provides a computer device, which includes a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to implement corresponding method flows or corresponding functions, thereby realizing the steps of the intelligent computing center load prediction method based on task element feature modulation in the above embodiments.
[0122] Example 4 Based on the same inventive concept, this invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the steps of the intelligent computing center load prediction method based on task element feature modulation in the above embodiments.
[0123] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0124] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting the load of a smart computing center based on task element feature modulation, characterized in that, The method includes: An adaptive wavelet packet decomposition method is used to perform multi-level decomposition of the total load sequence of the intelligent computing center, resulting in multiple sub-band component sequences. The dynamic weights of the subband component sequence in the training domain and the dynamic weights of the inference domain are determined based on the normalized mutual information between the subband component sequence and the features of each task element. A state-space model for benchmark prediction of subbands within the prediction interval is constructed based on the subband component sequences and their corresponding dynamic weights in the training and inference domains. The state-space model is used to perform baseline predictions for each sub-band within the prediction interval, resulting in a baseline prediction sequence for each sub-band within the prediction interval. Based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence, the baseline prediction sequence of each sub-band in the prediction interval is corrected to obtain the corrected baseline prediction sequence of each sub-band in the prediction interval. Perform inverse wavelet transform on the benchmark prediction sequence of each sub-band within the prediction interval after correction, and reconstruct the benchmark prediction sequence of each sub-band within the prediction interval into the time-domain load prediction result.
2. The method as described in claim 1, characterized in that, In the process of using the adaptive wavelet packet decomposition method to perform multi-level decomposition of the total load sequence of the intelligent computing center, when the current decomposition depth k satisfies When k is used as the maximum decomposition depth, where, For the first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of the nth frequency band node and the th Layer wavelet packet decomposition layer 1 The sum of the logarithmic energy entropy of each frequency band node For the first Layer wavelet packet decomposition layer 1 Logarithmic energy entropy of each frequency band node.
3. The method as described in claim 2, characterized in that, The first Layer wavelet packet decomposition layer 1 The logarithmic energy entropy of each frequency band node is as follows: In the above formula, For the first Layer wavelet packet decomposition layer 1 Number of discrete data points per frequency band node For the first Layer wavelet packet decomposition layer 1 The i-th discrete data point of a frequency band node It is a positive smoothing factor.
4. The method as described in claim 1, characterized in that, The task meta-features include at least one of the following: resource allocation features, training task features, inference task features, interconnection communication features, and scheduling event features.
5. The method as described in claim 4, characterized in that, The resource allocation features include the number of GPUs allocated, the number of CPU cores, the amount of video memory used, and the amount of RAM used; the training task features include the model parameter size, training status identifier, gradient synchronization event identifier, distributed parallel configuration, and checkpoint saving event; the inference task features include the request arrival rate, request queue length, input / output sequence length, model loading status, and changes in online service instances; the interconnection and communication features include the network throughput between nodes, the bandwidth utilization between chips, and the communication blocking flag.
6. The method as described in claim 1, characterized in that, The normalized mutual information between the sub-band component sequences and the meta-features of each task is as follows: In the above formula, Subband component sequence With task meta-features Normalized mutual information between them Subband component sequence With task meta-features Mutual information between them Subband component sequence Information entropy Task meta-features Information entropy.
7. The method as described in claim 6, characterized in that, The dynamic weights of the sub-band component sequences in the training domain and the dynamic weights in the inference domain are as follows: In the above formula, The training domain dynamic weights for the sub-band component sequences. The training domain association score is used for the sub-band component sequence. , The inference domain association score for the sub-band component sequence. , For the dynamic weights of the inference domain of the sub-band component sequences, The set of training domain features in the task meta-features. The first feature is used to aggregate weights. This is the set of inference domain features in the task meta-features. The aggregate weights are for the second feature.
8. The method as described in claim 7, characterized in that, The state-space model is as follows: In the above formula, For sub-band i at time i The hidden state vector. For sub-band i at time i Discrete state transition parameters, For sub-band i at time i Discrete input mapping parameters, For sub-band i at time i -1 hidden state vector For sub-band i at time i sequence of components, For sub-band i at time i The baseline forecast value, For sub-band i at time i The time-varying mapping parameters, Input the pass-through term for sub-band i.
9. The method as described in claim 8, characterized in that, The sub-band i at time i The discrete state transition parameters and discrete input mapping parameters are as follows: In the above formula, For sub-band i at time i The basic state transition matrix, For sub-band i at time i Time-varying input mapping parameters in continuous state space, For sub-band i at time i The discretization step size parameter, This is a discretization processing function.
10. The method as described in claim 9, characterized in that, The sub-band i at time i The time-varying input mapping parameters and subband i in the continuous state space at time... The time-varying mapping parameters are as follows: In the above formula, For a moment Gated generation function, Generate functions for input mapping. For the output mapping generation function, where the time... The gating generation function is as follows: The input mapping generation function and the output mapping generation function are as follows: , In the above formula, It is a non-linear activation function. For gating weights, For gated bias. It is a comprehensive feature vector composed of the current input, the local context features of the current input, and dynamically assigned weights. Let be the first projection matrix. For the first bias term, This is the second projection matrix. This is the second bias term.
11. The method as described in claim 10, characterized in that, The sub-band i at time i The discretization step size parameters are as follows: In the above formula, For activation function, This is the weight matrix. For the third bias term, This is the first positive adjustment coefficient corresponding to subband i. This is the second positive adjustment coefficient corresponding to subband i. For sub-band i at time i The dynamic weights of the training domain For sub-band i at time i Dynamic weights of the inference domain.
12. The method as described in claim 11, characterized in that, Before correcting the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval, the following steps are included: If subband i is at time i If the difference between the dynamic weights of the inference domain and the dynamic weights of the training domain is greater than or equal to 1, then subband i is classified as a high-frequency transient subband; otherwise, subband i is classified as a low-frequency steady-state subband.
13. The method as described in claim 12, characterized in that, For the low-frequency steady-state subband, the subband m at time t is determined by the following formula. Gain compensation coefficient: In the above formula, For sub-band m at time Gain compensation coefficient, As the lower limit of the calibration coefficient, To ensure a safe upper limit for the calibration coefficient, Let m be the linear mapping weights for subband m. For sub-band m at time steady-state scale-type feature input, The linear mapping bias for subband m For sub-band m at time Compensation training domain dynamic weights, For sub-band m at time The dynamic weights of the compensation inference domain; Subband m at time Baseline forecast Revised to .
14. The method as described in claim 13, characterized in that, For a high-frequency transient subband, the subband m at time t is determined by the following formula. Offset compensation term: In the above formula, For m at time Offset compensation term, , , These are the learnable linear mapping weights corresponding to traffic surge features, concurrent pressure features, and communication pulse features, respectively. For sub-band m at time Corresponding traffic surge characteristics For sub-band m at time The corresponding concurrent pressure characteristics, For sub-band m at time Corresponding communication pulse characteristics, This is the third bias term; Subband m at time Baseline forecast Revised to .
15. The method as described in claim 14, characterized in that, The sub-band m at time The process of obtaining the dynamic weights of the compensated training domain and the dynamic weights of the compensated inference domain includes: The dynamic weights of the subband component sequence compensation training domain and the dynamic weights of the compensation inference domain are determined based on the normalized mutual information between the subband component sequence and the meta-features of each compensation task.
16. The method as described in claim 15, characterized in that, The compensation task meta-features include: traffic surge features, concurrent pressure features, communication pulse features, and resource switching features.
17. The method as described in claim 16, characterized in that, The traffic surge characteristics include: active parameter scale, number of GPUs allocated, video memory usage, and number of training instances; the concurrency pressure characteristics include: request arrival rate, change in request arrival rate, queue length, and sudden increase in queue length; the communication pulse characteristics include: gradient synchronization flag, model loading event, pipeline switching event, and sudden increase in communication bandwidth; the resource switching characteristics include: load change event type.
18. A load forecasting device for intelligent computing centers based on task element feature modulation, characterized in that, The device includes: The decomposition module is used to perform multi-level decomposition of the total load sequence of the intelligent computing center using the adaptive wavelet packet decomposition method to obtain multiple sub-band component sequences. The determination module is used to determine the dynamic weights of the training domain and the dynamic weights of the inference domain of the sub-band component sequence based on the normalized mutual information between the sub-band component sequence and the features of each task element. The building module is used to construct a state-space model for benchmark prediction of subbands within the prediction interval based on the subband component sequence and its corresponding dynamic weights in the training and inference domains. The prediction module is used to perform baseline predictions for each sub-band within the prediction interval using the state space model, and obtain the baseline prediction sequence for each sub-band within the prediction interval. The correction module is used to correct the baseline prediction sequence of each sub-band within the prediction interval based on the dynamic weights of the training domain and the dynamic weights of the inference domain of each sub-band component sequence, so as to obtain the corrected baseline prediction sequence of each sub-band within the prediction interval. The analysis module is used to perform inverse wavelet transform on the benchmark prediction sequence of each sub-band within the prediction interval, and reconstruct the benchmark prediction sequence of each sub-band within the prediction interval into a time-domain load prediction result.
19. A computer device, characterized in that, include: One or more processors; The processor is used to execute one or more programs; When the one or more programs are executed by the one or more processors, the intelligent computing center load prediction method based on task element feature modulation as described in any one of claims 1 to 17 is implemented.
20. A computer-readable storage medium, characterized in that, It contains a computer program, which, when executed, implements the intelligent computing center load prediction method based on task element feature modulation as described in any one of claims 1 to 17.
