AI data center it load prediction method and system for heterogeneous cluster scenario

Abstract

The application provides an AI data center IT load prediction method and system for a heterogeneous cluster scene, collects cluster IT load sequences and workload state quantities; performs seasonal-trend decomposition on a target sequence of a training set to define a periodicity index by a daily seasonal component variance proportion; adaptively selects a calendar feature set, a workload feature set or a full feature set according to a comparison between the periodicity index and a preset threshold; adopts two-stage training of random mask type self-supervised pre-training, initialization of a backbone network by pre-training parameters and fine-tuning by time series Mixup enhancement to output multi-quantile load prediction; and calculates a widening coefficient by a non-increasing function of the periodicity index to post-process and widen a prediction interval. The application stably improves the accuracy without degradation on four types of clusters with a periodicity difference of two orders of magnitude, so that the interval coverage rate of a low periodicity cluster reaches a nominal level and a high periodicity cluster is not affected, thereby providing a reliable basis for scheduling planning of an AI data center connected to a power grid.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to an AI data center IT load forecasting method and system for heterogeneous cluster scenarios. Background Technology

[0002] The demand for artificial intelligence computing power has been surging in recent years, reshaping the global power supply and demand landscape. Several leading technology companies have recently announced large-scale expansion plans for AI data centers, with predictions suggesting that the total power consumption of such facilities will exceed 100 gigawatts by 2030. For power system operators, the grid connection of such massive, concentrated, and fluctuating loads requires high-confidence load forecasting, from hourly to day-ahead timeframes, to support grid planning, operational economics calculations, and the assessment of backup and ancillary services.

[0003] In the overall energy consumption of AIDC (AI Data Center), IT equipment load dominates. When the Power Usage Effectiveness (PUE) is typically around 1.12, IT equipment consumes approximately 89% of the total facility power, with the remaining 11% consumed in cooling, lighting, and auxiliary equipment. The cooling load largely varies with the heat dissipation of IT equipment and can be estimated using the IT load and PUE. Therefore, accurately predicting the IT load constitutes a core constraint in AIDC overall power forecasting.

[0004] Unlike conventional power loads, the IT load of AI data centers is driven by the flow of computing tasks (encompassing job submission, scheduling, and the arrival of inference calls), and is not dependent on temperature or people's daily routines. This renders the traditional "temperature overlay calendar" method of load forecasting ineffective. A particularly challenging aspect is the highly heterogeneous temporal behavior of AI computing tasks across different centers: clusters primarily handling long training tasks lasting several hours exhibit a nearly constant aggregate load with minimal intraday fluctuations; while inference clusters serving online user requests display loads highly synchronized with human schedules and exhibiting distinct daily cyclical characteristics. A single, fixed forecasting scheme is insufficient to accommodate these diverse cluster scenarios.

[0005] Existing solutions each have limitations, including: GPU and server power models at the device and node levels can only provide instantaneous power estimates and cannot provide multi-step rolling predictions; existing research on data center-grid interaction mostly treats the center load as an external input and does not address the prediction of IT load itself; deep learning predictions for data center cooling and power consumption mostly treat the entire center as a unified thermodynamic object and do not take into account the load evolution unique to AI clusters; and the few studies specifically for AI clusters mostly perform single-step recursive extrapolation in a single scenario, lacking both cross-scenario transfer and load characteristic perception and uncertainty probability characterization. In addition, the prediction range of quantile regression often has insufficient coverage under low-predictability conditions, and its direct use for scheduling will underestimate operational risks. Summary of the Invention

[0006] This invention provides an AI data center IT load prediction method and system for heterogeneous cluster scenarios, in order to solve the defects existing in the prior art.

[0007] In a first aspect, the present invention provides an AI data center IT load prediction method for heterogeneous cluster scenarios, comprising: Collect the IT load time series and workload status of the AI ​​data center computing cluster, preprocess the data and divide it into training set, validation set and test set; Seasonal-trend decomposition is performed on the target variable sequence of the training set to extract the daily seasonal component. The periodic index is obtained by the ratio of the variance of the daily seasonal component to the variance of the original sequence. The periodic index is compared with a preset upper threshold and a preset lower threshold. When the periodic index is higher than the preset upper threshold, the calendar feature set is selected; when it is lower than the preset lower threshold, the workload feature set is selected; when it is in between, the full feature set consisting of the union of the calendar feature set and the workload feature set is selected. The input of the prediction model is constructed using the selected feature set. A random mask is applied to the input window, and the backbone network is pre-trained in a self-supervised manner with the reconstruction error of the mask position as the loss to obtain the pre-trained parameters. The backbone network is initialized with pre-trained parameters, the training sample pairs are enhanced by temporal convex combination, fine-tuned with multi-quantile bouncing loss function, and the load prediction value under the preset quantile is output. Using the median quantile prediction value as the point prediction, the interval widening coefficient is calculated using a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound. The upper and lower quantile prediction values ​​are then widened proportionally to obtain the adjusted prediction interval.

[0008] According to the present invention, an AI data center IT load prediction method for heterogeneous cluster scenarios is provided, wherein the IT load time series uses IT equipment power or the number of active GPUs as the target variable; The workload status parameters are taken from at least one of the following: number of running jobs, job queue depth, number of active GPUs, job submission count, and job completion count; The statistics required for normalization are estimated only on the training set and then applied to the validation and test sets.

[0009] According to the present invention, an AI data center IT load prediction method for heterogeneous cluster scenarios is provided, wherein the seasonal-trend decomposition adopts STL decomposition, and the periodic index takes a value between 0 and 1. The physical basis for the periodicity index is that the longer the average service time of the operation, the more significantly the daily periodic component is weakened, and the smaller the corresponding periodicity index.

[0010] According to the present invention, an AI data center IT load prediction method for heterogeneous cluster scenarios is provided, wherein the calendar feature set includes the target variable itself, the first difference of the target variable, the moving mean and moving standard deviation of the target variable, and the periodic encoding of the hour and week. The workload feature set includes the target variable, the workload state variables and their respective first differences, and the sliding statistics of the target variable.

[0011] According to the present invention, an AI data center IT load prediction method for heterogeneous cluster scenarios is provided. In the self-supervised pre-training stage, each element of the mask matrix is ​​independent and takes a value with a preset mask rate as a parameter. The pre-training loss is the mean square reconstruction error of the mask position, excluding the predicted label.

[0012] According to the present invention, an AI data center IT load prediction method for heterogeneous cluster scenarios is provided. In the fine-tuning stage, the mixing coefficient of the convex combination enhancement takes the value of a preset distribution, and the multi-quantile bouncing loss function includes at least the sum of bouncing losses at the preset upper quantile, median quantile, and preset lower quantile.

[0013] According to the present invention, an AI data center IT load prediction method for heterogeneous cluster scenarios is provided, wherein the functional form of the interval widening coefficient includes: When the periodicity index is zero, it takes a preset maximum value; when the periodicity index is greater than or equal to a preset critical value, it takes 1.0; and it decreases linearly in the middle section. The adjustment is a post-processing operation that does not modify the point prediction; the periodic index is determined by performing a single seasonal-trend decomposition on the training set.

[0014] Secondly, the present invention also provides an AI data center IT load prediction system for heterogeneous cluster scenarios, comprising: The preprocessing module is used to collect the IT load time series and workload status of the AI ​​data center computing cluster, preprocess the data and divide it into training set, validation set and test set; The quantization module is used to perform seasonal-trend decomposition on the target variable sequence of the training set, extract the daily seasonal component, and obtain the periodic index by the ratio of the variance of the daily seasonal component to the variance of the original sequence. The comparison module is used to compare the periodic index with a preset upper threshold and a preset lower threshold. When the periodic index is higher than the preset upper threshold, the calendar feature set is selected; when it is lower than the preset lower threshold, the workload feature set is selected; when it is in between, the full feature set formed by the union of the calendar feature set and the workload feature set is selected. The input of the prediction model is constructed using the selected feature set. The configuration module is used to apply a random mask to the input window and perform self-supervised pre-training on the backbone network with the reconstruction error of the mask position as the loss to obtain pre-training parameters; The training module is used to initialize the backbone network with pre-trained parameters, perform temporal convex combination enhancement on training sample pairs, fine-tune it with the multi-quantile bouncing loss function, and output the load prediction value under the preset quantile. The adjustment module is used to use the median quantile prediction value as the point prediction, calculate the interval widening coefficient with a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound, and proportionally widen the upper and lower quantile prediction values ​​to obtain the adjusted prediction interval.

[0015] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the AI ​​data center IT load prediction method for heterogeneous cluster scenarios as described above.

[0016] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI ​​data center IT load prediction method for heterogeneous cluster scenarios as described above.

[0017] The AI ​​data center IT load prediction method and system for heterogeneous cluster scenarios provided by this invention have the following beneficial effects: 1. Use cyclical indices π It uniformly characterizes the daily cycle strength of IT load in heterogeneous clusters and provides its physical explanation through the spectral decay mechanism of queuing theory, so that feature selection, training settings and interval calibration can all be adapted to the scenario, overcoming the shortcomings of a single solution in taking into account heterogeneous clusters. 2. The two-step training consisting of masked self-supervised pre-training followed by time-series Mixup enhancement fine-tuning is complementary in the frequency domain and does not depend on the backbone structure. It stably improves accuracy on four types of clusters with a π phase difference of two orders of magnitude, without any configuration degradation. 3. By π The driven interval widening rule increases the empirical coverage of weak-periodic clusters at the 80% nominal level from about 72% to above 85%, stably exceeding the nominal value, while keeping the interval of strong-periodic clusters unchanged and without altering point predictions. The inference burden is negligible, providing reliable probabilistic predictions of coverage for power grid dispatch planning. 4. Relying on π The predictive applicability classification allows for the determination of the scheduling mode of a new grid-connected cluster using only a few weeks of data, without the need to build a predictive model in advance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the AI ​​data center IT load prediction method for heterogeneous cluster scenarios provided by the present invention; Figure 2 This is the second flowchart of the AI ​​data center IT load prediction method for heterogeneous cluster scenarios provided by the present invention. Figure 3 This is a schematic diagram of the two-stage training strategy structure provided by the present invention; Figure 4 This is a schematic diagram of the interval widening coefficient function curve driven by the periodic index provided by the present invention and the cluster values ​​of four types of embodiments; Figure 5 This is a schematic diagram of the prediction applicability partitioning based on periodic indices and the cluster prediction accuracy of four types of embodiments provided by the present invention; Figure 6 This is a schematic diagram of the structure of the AI ​​data center IT load prediction system for heterogeneous cluster scenarios provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] To address the problems existing in the current technology, this invention proposes a method using periodic indices. π This AI data center IT load prediction method uniformly characterizes the daily cycle strength of cluster IT load and adaptively sets input features, training methods, and prediction intervals accordingly, thereby compensating for the shortcomings of a single prediction scheme that cannot simultaneously handle heterogeneous clusters such as training, hybrid, and inference clusters.

[0022] Another objective of this invention is to address the technical problems existing in the prior art by providing a two-step training strategy that consists of masked self-supervised pre-training followed by temporal Mixup enhancement fine-tuning, and is independent of the specific backbone architecture. The two stages are complementary in the frequency domain, where pre-training converges the backbone network to the low-frequency skeleton of the target sequence, and Mixup applies perturbation to the high-frequency components, which is equivalent to selective regularization along the frequency dimension, thereby generally improving accuracy without degradation in heterogeneous cluster scenarios.

[0023] Another objective of this invention is to address the technical problems existing in the prior art and to provide a method for dealing with periodic indices. π The post-processing method for the prediction interval of the regulation expands the interval proportionally according to the predictability of the cluster itself, so that the experience coverage of the weak periodic cluster can reliably reach the nominal level, while keeping the interval of the strong periodic cluster unchanged, and the cost of point prediction and inference is almost zero.

[0024] Figure 1 This is one of the flowcharts illustrating the AI ​​data center IT load prediction method for heterogeneous cluster scenarios provided in this embodiment of the invention, such as... Figure 1 As shown, it includes: Step 100: Collect the IT load time series and workload status of the AI ​​data center computing cluster, preprocess the data and divide it into training set, validation set and test set; Step 200: Perform seasonal-trend decomposition on the target variable sequence of the training set, extract the daily seasonal components, and obtain the periodic index by the ratio of the variance of the daily seasonal components to the variance of the original sequence. Step 300: Compare the periodicity index with the preset upper threshold and the preset lower threshold. When the periodicity index is higher than the preset upper threshold, select the calendar feature set; when it is lower than the preset lower threshold, select the workload feature set; when it is between the two, select the full feature set formed by the union of the calendar feature set and the workload feature set. Construct the input of the prediction model with the selected feature set. Step 400: Apply a random mask to the input window, and perform self-supervised pre-training on the backbone network with the reconstruction error of the mask position as the loss to obtain the pre-training parameters; Step 500: Initialize the backbone network with pre-trained parameters, perform temporal convex combination enhancement on the training sample pairs, fine-tune it with the multi-quantile bouncing loss function, and output the load prediction value under the preset quantile. Step 600: Using the median quantile prediction value as the point prediction, calculate the interval widening coefficient using a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound, and widen the upper and lower quantile prediction values ​​proportionally to obtain the adjusted prediction interval.

[0025] Specifically, the embodiments of the present invention include the following steps: Step 1, Data Acquisition and Preprocessing: Collect the IT load time series and workload status of the AI ​​data center computing cluster. The IT load time series uses IT equipment power or the number of active GPUs as the target variable. Resample each series to a consistent time granularity and divide them into three subsets according to time sequence: training, validation, and testing. The statistics required for normalization are estimated only on the training subset and then applied to the validation and testing subsets. Step 2, Periodic Quantization Step: Perform seasonal-trend decomposition on the target variable sequence of the training set, extract the daily seasonal components, and calculate the periodic index. π The periodic index π Defined as the ratio of the variance of the daily seasonal component to the variance of the original sequence, it is used to quantify the strength of the daily periodicity of cluster IT load. Step 3, Scene-Oriented Feature Set Adaptive Configuration Step: This involves configuring the periodic index... π Compared with the preset upper and lower thresholds, π When the value exceeds the upper threshold, the calendar feature set is selected. π When the value is below the lower threshold, the workload feature set is selected. π When the value is between the upper and lower thresholds, the full feature set consisting of the union of the calendar feature set and the workload feature set is selected, and the input of the prediction model is constructed using the selected feature set. Step 4, two-stage training steps: In the first stage, a random mask is applied to the input window, and the reconstruction error of the mask position is used as the loss to perform self-supervised pre-training on the backbone network to obtain pre-training parameters; In the second stage, the above pre-training parameters are used to assign initial values ​​to the backbone network, temporal Mixup convex combination enhancement is applied to paired training samples, and fine-tuning is completed with the multi-quantile bouncing loss function so that the backbone network can give load predictions at each preset quantile. Step 5, Prediction Interval Adjustment Steps: Using the median quantile prediction value as the point prediction, based on the periodic index... π Calculate the interval widening factor w ( π The w ( π ) is about π A monotonically non-increasing function bounded by 1.0, with the upper and lower quantile predictions centered around the median. w ( π The prediction interval is widened proportionally to obtain the adjusted prediction interval, and the output point prediction is compared with the adjusted prediction interval.

[0026] In the above method, in step 2, the seasonal-trend decomposition uses STL decomposition, and the periodic index is calculated using the following formula:

[0027] In the formula, S daily It is the daily seasonal term obtained from STL decomposition. Y It is the original sequence of target variables. Var(·) table variance, π The value falls within [0,1], reflecting the proportion of the daily periodic component in the total variance. Periodic index π The physical basis can be given by queuing theory: abstracting a computing power cluster as an M / G / k queuing model, its job arrival rate has a daily periodic modulation, and the service link plays a low-pass filtering effect on this arrival rate, so the number of active service stations at a daily frequency... ω d The power spectral density on follows:

[0028] In the formula, α d Record the daily modulation amplitude of the arrival rate. E [ S Record the average service time for the task. ρ Record the service desk occupancy rate. Once... E [ S As the number of operations increases, the daily cycle component is significantly suppressed: training clusters that primarily perform long operations lasting several hours... πThe value approaches zero, while inference clusters that primarily handle requests at the second level... π Approaching 1. This pattern explains the cause of the predictability differences between heterogeneous clusters, and is also the basis of this invention. π This serves as the basis for unified regulation.

[0029] In the above method, step 3, the calendar feature set consists of the following: the target variable itself and its first difference, the moving mean and moving standard deviation of the target variable, plus the periodic encoding of hours and days of the week; the workload feature set includes the target variable, one or more of the status variables listed above (number of running jobs, job queue depth, number of active GPUs, number of submitted and completed jobs, etc.) and their first differences, plus the moving statistics of the target variable; the two are combined to obtain the full feature set. The upper threshold can be set to 50%, and the lower threshold can be set to 15%; by π The selected candidate feature set is ultimately validated using the loss on the validation subset.

[0030] In the above method, in the first stage of step 4, the input window... ( L To check the window length, F Apply a random mask to the feature number:

[0031] In the formula, the mask matrix M Each element in the code is independent of the others, and the mask rate is used to define the elements. r (For example, 0.3) is a parameter that follows a Bernoulli distribution, and ⊙ denotes the element-wise product. f θ Record the backbone network, Let the linear reconstruction output head be denoted. The pre-training loss is taken as the mean square reconstruction error at the masked position:

[0032] This stage does not require any predicted labels; its aim is to learn the relationships between features and their temporal coherence. After training, the backbone network parameters are... Keep it.

[0033] In the above method, in the second stage of step 4, using Assign initial values ​​to the backbone network, and take two samples within each batch to perform a convex combination:

[0034] In the formula, the mixed weights λ Taken from Beta ( α , α Distribution. The model applies to quantiles. τ Find the minimum loss of the bouncing ball at points ∈{0.1, 0.5, 0.9}:

[0035] In the formula Used for point prediction, Enclose 80% of the prediction interval. Hybrid hyperparameters. α The strength of the interpolation is determined by Bayesian optimization (sampling with a tree-structured Parzen estimator) based on the validation set loss, and can be optimized within the range of [0.005, 5.0]. α optimal Value and π The reflected cluster signal-to-noise ratio matches, that is π For situations with low noise and high noise levels, a smaller size is preferable. α (Weak enhancement), for cases with high signal-to-noise ratio, a larger value can be selected. α (Strongly hybrid), therefore, when targeting a new cluster, it can be... Given π and α Initialize values ​​to avoid exhaustive search. Fine-tuning phase to verify subset performance triggers early stopping.

[0036] The two-stage combination has a complementary effect: pre-training converges the backbone network to the low-frequency temporal skeleton of the target sequence, while Mixup stirs up the high-frequency part, equivalent to a frequency-oriented selective regularizer. Without pre-training, Mixup may overwhelm the temporal patterns unique to the workload, causing the accuracy of clusters with sufficient signal to decline; however, after pre-training, only the high-frequency residuals are subject to regularization constraints, and the two-stage combination yields positive gains in heterogeneous cluster scenarios.

[0037] In the above method, in step 5, the interval widening coefficient and the adjusted quantile are calculated using the following formula:

[0038]

[0039] w ( π ) π Decreasing: When π =0 (no periodicity) w =1.5, the interval is correspondingly expanded by 50%; when π ≥0.625 w =1.0, the interval remains unchanged. This adjustment is a post-processing operation and does not modify point predictions; the used π The value only needs to be selected right The training subset is obtained by performing a single STL decomposition, and the inference process introduces almost no additional computational cost. The two constants (1.5, 0.8) were obtained by fitting a uniform linear regression on several heterogeneous cluster datasets. They are generally conservative to ensure that weakly periodic clusters can consistently exceed the nominal coverage.

[0040] The above method may also include a step of predicting applicability classification: based on cyclical indices. π The cluster is divided into three segments: when π > 50%, its IT load is considered sufficient to support trusted day-ahead scheduling; when 15% ≤ π, its load is considered sufficient to support trusted day-ahead scheduling; when π ... π ≤50%, it is advisable to combine daily rolling correction with prior forecast; when π If the load is less than 15%, the classification must be based on short-term rolling forecasts (e.g., 4 hours). This classification only requires calculations from several weeks of historical load data. π It can be determined that no trained model is required, and it can be used for rapid pre-screening of new grid-connected clusters.

[0041] In the above method, the backbone network can be selected from three types: linear decomposition model (DLinear), inverted Transformer model (iTransformer), and fragmented Transformer model (PatchTST). The two-step training strategy does not depend on the specific structure of the backbone network.

[0042] The implementation scheme of this application is illustrated below with a specific example.

[0043] like Figure 2 As shown, the overall process of the method of the present invention is as follows: First, the IT load time series and workload state variables of the AI ​​data center computing cluster are collected and preprocessed; then, the periodic index is calculated on the training set through seasonal-trend decomposition. π Based on π The comparison results with the preset threshold are used to complete the adaptive selection of the feature set for the scene; then, the backbone network is trained in two steps, consisting of masked self-supervised pre-training followed by temporal Mixup enhancement fine-tuning cascade, to output multiquantile load prediction; finally, the prediction is... π Calculate the interval widening factor, perform post-processing adjustments on the prediction interval, and output the point prediction and the adjusted prediction interval. The following describes the specific implementation methods of each step using four implementation datasets.

[0044] This embodiment selects four real-world datasets that represent typical AI cluster architectures to implement the method of this invention: (a) MIT Supercloud academic training cluster, 448 V100 GPUs, 234 days of data, with the target variable being IT power (kilowatts) and a periodicity index. π =1.85%; (b) Helios Saturn industrial training cluster, 2072 GPUs, 204 days of data, target variable is the number of active GPUs. π =4.50%; (c) Alibaba PAI hybrid machine learning platform, 10412 A100 / H800 GPUs, 184 days of data, target variable is the number of active GPUs. π=22.31%; (d) Inference service cluster scenario, 256 V100 GPUs, 120 days of data, the target variable is IT power (kilowatts). π =82.89%, reconstructed from publicly available LLM inference request trajectories combined with measured GPU power models. The periodicity exponents of the four clusters span two orders of magnitude, constituting a heterogeneous cluster scenario.

[0045] Step 1: Data Acquisition and Preprocessing. Collect IT load sequences and workload status metrics (number of running jobs, job queue depth, number of active GPUs, job submission and completion counts, etc.) for each cluster, and resample to a 15-minute granularity. Divide the data into training, validation, and test subsets according to time order at 70% / 15% / 15%. Standardize the target variable to zero mean and unit variance; standardized statistics are estimated only on the training subset. The lookback length L = 96 steps (24 hours), and the prediction step size H is set to 16 steps (4 hours) and 96 steps (24 hours). Sliding window samples are generated independently within each subset, with no window crossing subset boundaries. Multi-resolution comparison experiments show that the 15-minute resolution has the lowest root mean square error in both the 4-hour and 24-hour time domains, representing the optimal trade-off between accuracy and computational cost.

[0046] In practice, IT load time series data can be derived from existing data center infrastructure monitoring systems, such as the total power consumption reported by the server baseboard management controller, the branch power metering of the rack power distribution unit, or calculated from the number of active GPUs recorded by the cluster scheduling system. Workload status parameters are derived from the cluster job scheduler's operation logs, including the number of running jobs, the number of queued jobs, and the count of job submission and completion events. Before aggregation, sequences from different data sources need to be timestamped and then resampled uniformly to avoid introducing spurious correlations due to cross-source time misalignment. Missing values ​​during the data collection process are filled using linear interpolation, and time periods covered by long gaps are removed when constructing training samples.

[0047] Step 2, Periodicity Quantization. The target sequences in the training sets of each cluster are decomposed using STL to calculate the periodicity index. The results are shown above and are consistent with the theoretical ranking obtained by substituting the average service time of each cluster's jobs (MIT: 10.55 hours, Helios: 1.46 hours, Alibaba: approximately 0.5 hours, and inference scenario: approximately 3 seconds) into the queuing theory spectrum decay formula, thus verifying the physical basis of the periodicity index.

[0048] STL decomposition uses the number of steps per day as the seasonal periodicity parameter (96 steps at 15-minute resolution) and employs robust fitting to reduce the impact of occasional power spikes on component estimation. The variance of the daily seasonal components obtained from the decomposition is calculated point by point, and divided by the variance of the original sequence to obtain the periodicity index. In practice, the periodicity index is not sensitive to data length; stable estimates can be obtained from several weeks of historical data. Furthermore, its value remains consistent across the 4-hour and 24-hour prediction time domains, indicating that the index characterizes the inherent properties of cluster load rather than being a product of specific prediction settings.

[0049] Step 3, adaptive configuration of feature sets based on the scenario. According to the π threshold rule: Inference service cluster ( π =82.89%>50%) selected the calendar feature set; Alibaba platform (15%≤ π =22.31%≤50%) use the full feature set; MIT and Helios clusters ( π <15%) The workload feature set is selected. Each candidate feature set is finally confirmed by validation set loss.

[0050] In this embodiment, the calendar feature set has 8 dimensions, including the target variable itself, the first difference of the target variable, the 4-hour rolling mean, the 4-hour rolling standard deviation, and the sine and cosine periodic codes of the hour and week numbers; the workload feature set, based on the target variable and its moving statistics, further introduces state variables such as the number of running jobs, job queue depth, number of active GPUs, and the count of submitted and completed jobs, as well as their first differences; if the target variable itself is the number of active GPUs, then duplicate columns are removed; the full feature set is the union of the above two.

[0051] Step 4, Two-Phase Training. The structure of the two-phase training strategy is as follows: Figure 3 As shown. The backbone network is selected from DLinear, iTransformer, and PatchTST based on the verification loss. The first step is to set the mask ratio. r =0.3, perform 30 rounds of masked self-supervised pre-training on the backbone; the second step is to perform temporal mixup fine-tuning after initialization with pre-trained weights; α The optimal algorithm is selected using Bayesian optimization on the interval [0.005, 5.0]. The optimizer is AdamW, the batch size is 128, the gradient cutoff threshold is set to 1.0, the learning rate is scheduled according to the OneCycle strategy, the random seed is fixed, and early stopping is triggered by the validation subset.

[0052] The results of the validation set loss optimization are as follows: In the 24-hour time domain, the MIT cluster uses a linear decomposition model with calendar features, the Helios cluster uses an inverted Transformer with workload features, the Alibaba platform uses a linear decomposition model with full features, and the inverted Transformer with calendar features is used for inference scenarios. π The direction given by the threshold rule is consistent, verifying that... π The effectiveness of a two-tiered configuration strategy—first narrowing the search space, then determining the outcome based on verification losses—is demonstrated. (Mixup hyperparameters) α The optimization results also show a pattern: the optimal clusters with low periodicity and high noise are the best. α Much less than 1 (corresponding to weak enhancement); optimal for high signal-to-noise ratio clusters α It can exceed 1 (corresponding to strong mixing), which is exactly the same as π The signal-to-noise ratio of the cluster is consistent.

[0053] Step 5, Prediction Interval Adjustment. The curve shape of the interval widening coefficient function and the value positions of the clusters in each embodiment are as follows: Figure 4 As shown. According to the definition of the widening coefficient. w ( π =max(1.0, 1.5) 0.8 π The values ​​were calculated for each cluster: MIT was 1.485, Helios was 1.464, Alibaba was 1.322, and the inference scenario was 1.000 (i.e., the interval remained unchanged). Based on this, the 80% prediction interval of each cluster was proportionally increased.

[0054] The sharpness cost of widening the interval is bounded: Measured by the average interval width as a percentage of the target value range after normalization, the MIT cluster widened from 21.9% to 32.6% in the 24-hour time domain, the Helios cluster widened from 18.1% to 26.5%, and the Alibaba platform widened from 10.8% to 14.2%. The inference scenario remained unchanged. After the adjustment, the interval width of all scenarios was less than 33%, which is still of reference value for scheduling decisions.

[0055] The results of the implementation are as follows.

[0056] (1) Point prediction accuracy: In the 24-hour prediction time domain, the method of this invention relative to the best benchmark of each cluster (covering ARIMA, persistence method, support vector regression, LSTM, LightGBM and 9 combinations of "modern backbone × feature set"): MIT cluster determination coefficient RThe mean square error increased from 0.323 to 0.329 for Helios, from 0.922 to 0.930, with the root mean square error decreasing from 149 GPUs to 141 and the symmetric mean absolute percentage error narrowing from 30.3% to 19.9%. For Alibaba, the increase was from 0.590 to 0.595; and for inference scenarios, from 0.873 to 0.876. In 72 comparisons with the same backbone across 9 backbone-feature combinations, 4 datasets, and 2 time domains, the proposed strategy improved 57 of these comparisons. R ² Improved, the rest remained unchanged, and no one deteriorated.

[0057] (2) Ablation analysis: Using either stage alone will result in degradation in some scenarios (e.g., when only Mixup is applied, Helios's...). R ² Decline of 0.020); After the two are combined, all four clusters turn to positive gains (MIT +0.006, Helios +0.008, Alibaba +0.005, inference scenario +0.005), and their values ​​exceed the sum of the gains of the two stages acting alone, confirming that the complementary effect is constructive.

[0058] (3) Interval coverage: Applying coverage at 80% nominal coverage π After the adjustment, the 24-hour experience coverage increased from 71.8% to 87.4% for MIT and from 74.0% to 85.0% for Alibaba, both steadily exceeding the nominal value; the inference scenario range remained unchanged (coverage 78.7%). The adjustment model remained stable in both the 4-hour and 24-hour time domains. The adjusted range width (percentage normalized to the target value range) was below 33% in all scenarios, indicating a manageable cost to sharpness.

[0059] (4) Applicability classification: such as Figure 5 As shown, the four clusters are arranged in order to π The data falls into three segments, which aligns with the measured order of predictability, confirming the following rule: π Those with over 50% can carry reliable day-ahead plans; those with 15% to 50% should use day-ahead forecasts with intraday adjustments; and those with less than 15% must use 4-hour rolling short-term forecasts.

[0060] In the 4-hour prediction time domain, the method of this invention is no less than the best benchmark on all clusters: the coefficient of determination reaches 0.786 for the MIT cluster, 0.979 for the Helios cluster, 0.731 for the Alibaba platform, and is on par with 0.879 for the inference scenario. Moreover, the accuracy of each method in the 4-hour time domain is generally higher than that in the 24-hour time domain, which is consistent with the rule that the shorter the prediction time domain, the stronger the predictability. This shows that the gain of the training strategy of this invention is more obvious in the longer and more difficult prediction time domain.

[0061] In other implementations: the periodicity index can also be calculated in the frequency domain, i.e., power spectrum estimation is performed on the target variable sequence, and the strength of daily periodicity is characterized by the proportion of energy in the frequency band near the daily frequency to the total energy; the upper and lower thresholds can be adjusted within the range of 40%~60% and 10%~20% based on the periodicity index distribution of the access cluster group; the output quantile level can be selected according to the required confidence level, for example, three levels of 0.05, 0.5 and 0.95 to form a 90% prediction interval; in addition to the total power of IT equipment and the number of active GPUs, the target variable can also be a meter with equivalent physical meaning, such as rack-level power or power supply branch power; the time resolution can be selected within the range of 5 minutes to 1 hour, and multi-resolution comparison shows that 15-minute resolution is the optimal compromise between prediction accuracy and computational load; the method of the present invention can be deployed in data center energy management systems, campus-level edge gateways or cloud prediction services, and interface with existing monitoring systems in the form of software modules, realizing the load prediction function according to the steps described in the claims.

[0062] The following describes the AI ​​data center IT load prediction system for heterogeneous cluster scenarios provided by this invention. The AI ​​data center IT load prediction system for heterogeneous cluster scenarios described below can be referred to in correspondence with the AI ​​data center IT load prediction method for heterogeneous cluster scenarios described above.

[0063] Figure 6 This is a schematic diagram of the structure of an AI data center IT load prediction system for heterogeneous cluster scenarios provided in an embodiment of the present invention, as shown below. Figure 6 As shown, it includes: a preprocessing module 61, a quantization module 62, a comparison module 63, a configuration module 64, a training module 65, and an adjustment module 66, wherein: Preprocessing module 61 is used to collect the IT load time series and workload status variables of the AI ​​data center computing cluster, preprocess the data, and divide it into training set, validation set, and test set; quantization module 62 is used to perform seasonal-trend decomposition on the target variable sequence of the training set, extract the daily seasonal component, and obtain the periodicity index by the ratio of the variance of the daily seasonal component to the variance of the original sequence; comparison module 63 is used to compare the periodicity index with a preset upper threshold and a preset lower threshold. When the periodicity index is higher than the preset upper threshold, the calendar feature set is selected; when it is lower than the preset lower threshold, the workload feature set is selected; when it is between the two, the full value formed by the union of the calendar feature set and the workload feature set is selected. The feature set is used to construct the input of the prediction model; the configuration module 64 is used to apply a random mask to the input window, and to perform self-supervised pre-training on the backbone network with the reconstruction error of the mask position as the loss to obtain the pre-training parameters; the training module 65 is used to initialize the backbone network with the pre-training parameters, to perform temporal convex combination enhancement on the training sample pairs, to fine-tune it with the multi-quantile bouncing loss function, and to output the load prediction value under the preset quantile; the adjustment module 66 is used to use the median quantile prediction value as the point prediction, to calculate the interval widening coefficient with a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound, and to proportionally widen the upper and lower quantile prediction values ​​to obtain the adjusted prediction interval.

[0064] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740. The processor 710, communication interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute an AI data center IT load prediction method for heterogeneous cluster scenarios. This method includes: collecting the IT load time series and workload status variables of the AI ​​data center computing cluster; preprocessing the data and dividing it into training, validation, and test sets; performing seasonal-trend decomposition on the target variable sequence of the training set to extract daily seasonal components; obtaining a periodic index by the ratio of the variance of the daily seasonal components to the variance of the original sequence; comparing the periodic index with a preset upper threshold and a preset lower threshold; selecting a calendar feature set when the periodic index is higher than the preset upper threshold and a workload feature set when it is lower than the preset lower threshold; and so on. The system uses a full feature set, formed by the union of the calendar feature set and the workload feature set, as the input to the prediction model. A random mask is applied to the input window, and the reconstruction error at the mask position is used as the loss for self-supervised pre-training of the backbone network to obtain pre-training parameters. The backbone network is initialized with these pre-training parameters, and temporal convex combination enhancement is performed on the training sample pairs. Fine-tuning is done using a multi-quantile bouncing loss function, and the load prediction value at the preset quantile is output. The median quantile prediction value is used as the point prediction, and an interval widening coefficient is calculated using a function that is monotonically non-increasing with respect to a periodic exponent and has a preset lower bound. The upper and lower quantile prediction values ​​are then proportionally widened to obtain the adjusted prediction interval.

[0065] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the above-described methods for predicting the IT load of AI data centers in heterogeneous cluster scenarios. This method includes: collecting the IT load time series and workload state variables of the AI ​​data center computing cluster; preprocessing the data and dividing it into a training set, a validation set, and a test set; performing seasonal-trend decomposition on the target variable sequence of the training set, extracting daily seasonal components, and obtaining a periodic index by the ratio of the variance of the daily seasonal components to the variance of the original sequence; comparing the periodic index with a preset upper threshold and a preset lower threshold; selecting a calendar feature set when the periodic index is higher than the preset upper threshold; and lowering the lower threshold. When the threshold value is set below the preset threshold, the workload feature set is selected. When the threshold value is between the two, the full feature set consisting of the union of the calendar feature set and the workload feature set is selected. The input of the prediction model is constructed using the selected feature set. A random mask is applied to the input window, and the reconstruction error of the mask position is used as the loss to perform self-supervised pre-training on the backbone network to obtain the pre-training parameters. The backbone network is initialized with the pre-training parameters, and the training sample pairs are enhanced by temporal convex combination. The multi-quantile bouncing loss function is used for fine-tuning, and the load prediction value at the preset quantile is output. The median quantile prediction value is used as the point prediction value. The interval widening coefficient is calculated using a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound. The upper and lower quantile prediction values ​​are widened proportionally to obtain the adjusted prediction interval.

[0067] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0069] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting IT load in AI data centers for heterogeneous cluster scenarios, characterized in that, include: Collect the IT load time series and workload status of the AI ​​data center computing cluster, preprocess the data and divide it into training set, validation set and test set; Seasonal-trend decomposition is performed on the target variable sequence of the training set to extract the daily seasonal component. The periodic index is obtained by the ratio of the variance of the daily seasonal component to the variance of the original sequence. The periodic index is compared with a preset upper threshold and a preset lower threshold. When the periodic index is higher than the preset upper threshold, the calendar feature set is selected; when it is lower than the preset lower threshold, the workload feature set is selected; when it is in between, the full feature set consisting of the union of the calendar feature set and the workload feature set is selected. The input of the prediction model is constructed using the selected feature set. A random mask is applied to the input window, and the reconstruction error of the mask position is used as the loss to perform self-supervised pre-training on the backbone network to obtain the pre-training parameters; The backbone network is initialized with pre-trained parameters, the training sample pairs are enhanced by temporal convex combination, fine-tuned with multi-quantile bouncing loss function, and the load prediction value under the preset quantile is output. Using the median quantile prediction value as the point prediction, the interval widening coefficient is calculated using a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound. The upper and lower quantile prediction values ​​are then widened proportionally to obtain the adjusted prediction interval.

2. The AI ​​data center IT load prediction method for heterogeneous cluster scenarios according to claim 1, characterized in that, The IT load time series uses IT equipment power or the number of active GPUs as the target variable; The workload status parameters are taken from at least one of the following: number of running jobs, job queue depth, number of active GPUs, job submission count, and job completion count; The statistics required for normalization are estimated only on the training set and then applied to the validation and test sets.

3. The AI ​​data center IT load prediction method for heterogeneous cluster scenarios according to claim 1, characterized in that, The seasonal-trend decomposition uses STL decomposition, and the periodic index takes values ​​between 0 and 1; The physical basis for the periodicity index is that the longer the average service time of the operation, the more significantly the daily periodic component is weakened, and the smaller the corresponding periodicity index.

4. The AI ​​data center IT load prediction method for heterogeneous cluster scenarios according to claim 1, characterized in that, The calendar feature set includes the target variable itself, the first difference of the target variable, the moving mean and moving standard deviation of the target variable, and the periodic encoding of hours and days of the week; The workload feature set includes the target variable, the workload state variables and their respective first differences, and the sliding statistics of the target variable.

5. The AI ​​data center IT load prediction method for heterogeneous cluster scenarios according to claim 1, characterized in that, In the self-supervised pre-training stage, each element of the mask matrix is ​​independent and takes the value of a preset mask rate as a parameter. The pre-training loss is the mean square reconstruction error of the mask position, excluding the predicted label.

6. The AI ​​data center IT load prediction method for heterogeneous cluster scenarios according to claim 1, characterized in that, During the fine-tuning stage, the mixing coefficients of the convex combination enhancement are taken from a preset distribution, and the multi-quantile bouncing loss function includes at least the sum of bouncing losses at the preset upper quantile, median quantile, and preset lower quantile.

7. The AI ​​data center IT load prediction method for heterogeneous cluster scenarios according to claim 1, characterized in that, The functional form of the interval widening coefficient includes: When the periodicity index is zero, it takes a preset maximum value; when the periodicity index is greater than or equal to a preset critical value, it takes 1.0; and it decreases linearly in the middle section. The adjustment is a post-processing operation that does not modify the point prediction; the periodic index is determined by performing a single seasonal-trend decomposition on the training set.

8. An AI data center IT load forecasting system for heterogeneous cluster scenarios, characterized in that, include: The preprocessing module is used to collect the IT load time series and workload status of the AI ​​data center computing cluster, preprocess the data and divide it into training set, validation set and test set; The quantization module is used to perform seasonal-trend decomposition on the target variable sequence of the training set, extract the daily seasonal component, and obtain the periodic index by the ratio of the variance of the daily seasonal component to the variance of the original sequence. The comparison module is used to compare the periodic index with a preset upper threshold and a preset lower threshold. When the periodic index is higher than the preset upper threshold, the calendar feature set is selected; when it is lower than the preset lower threshold, the workload feature set is selected; when it is in between, the full feature set formed by the union of the calendar feature set and the workload feature set is selected. The input of the prediction model is constructed using the selected feature set. The configuration module is used to apply a random mask to the input window and perform self-supervised pre-training on the backbone network with the reconstruction error of the mask position as the loss to obtain pre-training parameters; The training module is used to initialize the backbone network with pre-trained parameters, perform temporal convex combination enhancement on training sample pairs, fine-tune it with the multi-quantile bouncing loss function, and output the load prediction value under the preset quantile. The adjustment module is used to use the median quantile prediction value as the point prediction, calculate the interval widening coefficient with a function that is monotonically non-increasing with respect to the periodic exponent and has a preset lower bound, and proportionally widen the upper and lower quantile prediction values ​​to obtain the adjusted prediction interval.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the AI ​​data center IT load prediction method for heterogeneous cluster scenarios as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the AI ​​data center IT load prediction method for heterogeneous cluster scenarios as described in any one of claims 1 to 7.

Images