HPC job power consumption prediction method and system based on online incremental learning

Abstract

The application provides an HPC job power consumption prediction method and system based on online incremental learning, relates to the field of high-performance computing, initializes a prediction model according to offline job power consumption data, obtains newly sampled HPC job power consumption data in real time to form an input data stream, processes the input data stream by using the prediction model to output a power consumption prediction value at a future time, updates the model by using online incremental learning based on a model update timing determination method during processing of the input data stream, the model update timing determination method is that when new data is received, the deviation distance distribution change between a true value and a prediction value is monitored to obtain a model prediction performance decline score, whether the prediction model needs to be subjected to incremental learning is determined based on a preset incremental learning score threshold, and when the model prediction performance declines, the original model parameters are optimized and updated to meet the requirement of predicting the power consumption value at the future time, and the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of high-performance computing (HPC), and particularly relates to an HPC job power consumption prediction method and system based on online incremental learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] In the field of high-performance computing (HPC), power consumption prediction has always been an important research direction. HPC systems are usually composed of thousands of processors and storage devices, with highly parallel, complex computing and communication modes, so their power consumption is often very large; accurate prediction of the power consumption of HPC jobs can provide valuable reference information for system administrators and application developers to optimize the performance and energy efficiency of HPC jobs; in addition, in some application scenarios that require fine control of HPC job power consumption, such as energy management of supercomputing centers, power consumption constraints in cloud computing environments, power consumption prediction is crucial.

[0004] With the rapid development of machine learning and deep learning, machine learning and deep learning can analyze data and train prediction models through a large amount of historical data, improving the accuracy and precision of prediction; in machine learning methods, support vector machine (SVM), neural network (ANN), decision tree (DT), random forest (RF), etc. are widely used in HPC power consumption prediction; among them, SVM is a regression method that can improve the accuracy of prediction by adjusting the kernel function and the penalty factor; ANN is a multi-layer feedforward neural network based on backpropagation algorithm, which optimizes the prediction result by adjusting the model structure and weight value; DT is a classification and regression method, which classifies and predicts through tree structure power consumption data; RF is an ensemble learning method that predicts through multiple decision trees, which can improve the accuracy of prediction; the above methods have certain improvement in prediction accuracy compared with traditional statistical methods, but also have some disadvantages; for example, the time complexity of SVM method is high, and the processing effect of high-dimensional data is poor; ANN method needs a large amount of training data and time to train the network, and is easy to fall into local optimal solution; DT method is prone to overfitting and is not suitable for continuous variable prediction; RF method needs a lot of computing resources and time cost to build multiple decision trees, and is sensitive to noise and outliers.

[0005] With the continuous development of deep learning, its prediction accuracy for time series data has far surpassed that of traditional machine learning methods. Deep learning can mine latent features from data without requiring manual feature design. Some researchers have applied deep learning methods such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory Networks (LSTMs) to HPC power consumption prediction. CNNs extract features from power consumption data through convolution and pooling operations, and make predictions using multi-layer convolutional and fully connected layer structures. RNNs predict power consumption by learning the relationships between time series data; they can handle time-series and long-form data, but require higher quality power consumption data. Furthermore, Kandanond used different prediction methods: Autoregressive Integrated Moving Average (ARIMA), Artificial Neural Network (ANN), and Multiple Linear Regression (MLR) to predict energy consumption. Taheri S, Zhou Z, and Chen F et al. focused on coarse-grained and fine-grained energy prediction methods based on the different time granularities of time-series data. Coarse-grained energy prediction methods use global data (e.g., the overall energy consumption of a data center) to predict device energy consumption, which is efficient but less accurate. In contrast, fine-grained energy prediction methods predict the power consumption of each device individually, taking into account details such as device characteristics, workload, and environmental factors, resulting in higher accuracy but greater time complexity. They also analyzed the noise patterns of the sample data and proposed a method for power data denoising, using recursive autoencoders (RAE) and autoencoders (AE) to predict short-term and long-term data, respectively.

[0006] While machine learning and deep learning methods have made significant progress in HPC power consumption prediction compared to traditional methods, enabling the processing of large-scale power consumption data, the training and prediction of deep neural network models rely on the fundamental assumption that the data distribution exhibits static and constant characteristics. In practical power consumption prediction, most power consumption data is non-stationary. This instability stems from factors such as varying task loads, power supply noise interference, temperature effects, and adjustments to system power management strategies. These factors cause the probability distribution of power consumption data to change over time, leading to concept drift, which affects the model's predictive ability. Therefore, traditional offline models cannot accurately predict power consumption data with concept drift. Furthermore, traditional offline models require a large amount of data for offline training, consuming significant time and computational resources. During this time, new data may arrive, but the model has not yet learned from this new data—a situation known as the "cold start" problem. Therefore, even if an established offline model has a high generalization ability, it cannot effectively integrate old and new data as the sample size increases, leading to a decline in predictive ability. SUMMARY

[0007] Based on the time-varying characteristics of the job power consumption data, the traditional offline model needs to be calibrated, and it is difficult to adapt to the prediction of time-varying data, in order to overcome the shortcomings of the prior art, the present application provides a HPC job power consumption prediction method and system based on online incremental learning, when the model prediction performance decreases, the original model parameters are optimized and updated to meet the demand of predicting future power consumption value, and the prediction accuracy is improved.

[0008] To achieve the above object, one or more embodiments of the present application provide the following technical solutions:

[0009] The present application provides a HPC job power consumption prediction method based on online incremental learning.

[0010] The HPC job power consumption prediction method based on online incremental learning comprises:

[0011] According to the offline HPC job power consumption data, the prediction model is initialized;

[0012] Real-time acquisition of newly sampled HPC job power consumption data, and forming an input data stream;

[0013] The prediction model processes the input data stream to generate and output the power consumption prediction value at the future time;

[0014] In the process of processing the input data stream, the model is updated using online incremental learning based on the model update timing determination method;

[0015] Wherein, the model update timing determination method is to monitor the deviation distance distribution change between the true value and the prediction value when receiving new data, to obtain the model prediction performance decline score, and to determine whether the prediction model needs to be incrementally learned based on the preset incremental learning score threshold.

[0016] Further, the initialization of the prediction model is specifically:

[0017] Based on the offline HPC job power consumption data, the data samples are constructed using sliding window, and the prediction model is trained offline through the constructed data samples to obtain the trained prediction model.

[0018] Further, the prediction model is constructed based on neural network, and the data stream is input to output the prediction of the future time power consumption value.

[0019] Further, the specific processing steps when receiving new data are:

[0020] The prediction model is used to predict the new data to obtain the prediction value;

[0021] According to the deviation distance of the predicted value and the true value, a model prediction performance decline score is calculated;

[0022] The model prediction performance decline score is compared with an incremental learning score threshold value, and when the incremental learning score threshold value is exceeded, the new data is learned as an incremental sample to update the parameters of the prediction model.

[0023] Further, the model prediction performance decline score is a method of constructing a sliding window, which converts the deviation distance of the predicted value and the true value into a unified judgment standard in real time, that is, the model prediction performance decline score, and the greater the score, the more serious the model performance decline, and vice versa.

[0024] Further, the method of constructing a sliding window is specifically:

[0025] Two sliding data windows with lengths of large and small sliding windows W and W' are used to cache the deviation distance before t time and the deviation distance at t time, and the performance decline score of the prediction model at t time is obtained from the complementary probability of the right tail function of the standard Gaussian distribution.

[0026] Further, the setting of the incremental learning score threshold value divides the model prediction performance decline score into multiple equal parts, and the best model update time, that is, the incremental learning score threshold value, is found through the two evaluation indexes of MAE and MAPE.

[0027] The second aspect of the present application provides an HPC job power consumption prediction system based on online incremental learning.

[0028] The HPC job power consumption prediction system based on online incremental learning includes an initialization module, a data acquisition module, a power consumption prediction module and an online learning module.

[0029] The initialization module is configured to initialize the prediction model according to offline HPC job power consumption data.

[0030] The data acquisition module is configured to acquire real-time sampled HPC job power consumption data and form an input data stream.

[0031] The power consumption prediction module is configured to process the input data stream using the prediction model to generate and output the power consumption prediction value at the future time.

[0032] The online learning module is configured to use online incremental learning to update the model based on the model update time determination method during processing of the input data stream.

[0033] The model updating occasion judgment method is based on a preset incremental learning score threshold to determine whether the prediction model needs to be subjected to incremental learning by monitoring the deviation distance distribution change between the true value and the predicted value and obtaining a model prediction performance decline score when new data is received.

[0034] The third aspect of the present application provides a computer readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the HPC job power consumption prediction method based on online incremental learning according to the first aspect of the present application.

[0035] The fourth aspect of the present application provides an electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, wherein the processor implements the steps in the HPC job power consumption prediction method based on online incremental learning according to the first aspect of the present application when executing the program.

[0036] The above one or more technical solutions have the following beneficial effects:

[0037] Based on the defects of the traditional model, the HPC job power consumption prediction method and system based on online incremental learning provided by the present application have the following advantages:

[0038] (1) Strong adaptability: the online incremental learning framework model can quickly adapt to the change of time series data, so that accurate prediction results can be quickly generated; for power consumption prediction tasks that need to handle data with "concept drift" characteristics, online incremental learning can better utilize the change of data and improve the accuracy of the model.

[0039] (2) Less data sampling: since online learning can quickly learn knowledge from real-time data, it reduces the number of data samples and the time cost of data preprocessing to a certain extent, improves the efficiency of the model, and solves the "cold start" problem of traditional offline models.

[0040] (3) Real-time adjustment of model parameters: online incremental learning can quickly adjust model parameters to improve model accuracy; this is a very critical point for power consumption prediction tasks, which greatly improves the accuracy of prediction.

[0041] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0042] The accompanying drawings, which form a part of the present application, are used to provide further understanding of the present application, and the illustrative embodiments of the present application and their description are used to explain the present application, and do not constitute improper limitations on the present application.

[0043] Figure 1Flow chart of the method of the first embodiment.

[0044] Figure 2 Framework chart of the job power consumption prediction of the first embodiment.

[0045] Figure 3 Schematic diagram of the prediction model update timing determination method of the first embodiment.

[0046] Figure 4 Schematic diagram of the right tail function of the first embodiment.

[0047] Figure 5 MAPE curve diagram of the online incremental update model under different thresholds of the first embodiment.

[0048] Figure 6 Time cost diagram of the HPC job power consumption prediction of the first embodiment.

[0049] Figure 7 Flow chart of the offline training of the first embodiment.

[0050] Figure 8 Flow chart of the online training of the first embodiment. DETAILED DESCRIPTION

[0051] It should be noted that the following detailed description is illustrative only, and is intended to provide further description in order to provide a fuller enabling teaching of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0052] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, devices, components and / or combinations thereof, but do not preclude the presence or addition of one or more other features, steps, operations, devices, components and / or combinations thereof.

[0053] The experimental data is derived from the server running jobs in Jinan Supercomputing Center. For job power consumption prediction, a prediction model is constructed using a neural network; on the basis of the traditional neural network, the present application proposes an online incremental learning job power consumption prediction framework, which has wide applicability and can be applied to various traditional machine learning and deep learning models to improve the prediction accuracy.

[0054] Online incremental learning is similar to the sliding time window, constantly rolling learning, ensure fixed prediction length; when the incremental sample is added, first error judgment, if the error exceeds the threshold, the sample will be incremental learning. In error judgment, if the conventional judgment method is used, for example, assuming that the average error of all training samples is set as the error threshold e k , when |e| k , incremental learning is carried out. This judgment method has certain disadvantages: when the data dimension is inconsistent, it is difficult to set the threshold; therefore, the application also proposes a method for determining the model update time based on the model prediction performance decline score, which realizes the selection and update of the model parameters, reduces the number of model parameter updates, and improves the learning efficiency of the model.

[0055] Embodiment one

[0056] In one or more embodiments, an HPC job power consumption prediction method based on online incremental learning is disclosed, as shown in Figure 1 , comprising the following steps:

[0057] Step S1: initializing the prediction model according to the offline HPC job power consumption data.

[0058] Step S2: real-time acquisition of newly sampled HPC job power consumption data, and forming an input data stream.

[0059] Step S3: the prediction model processes the input data stream to generate and output the power consumption prediction value at the future time.

[0060] Step S4: during the processing of the input data stream, based on the model update time determination method, using online incremental learning to update the model;

[0061] Wherein, the model update time determination method is to monitor the deviation distance distribution change between the true value and the prediction value when receiving new data, to obtain the model prediction performance decline score, and to determine whether the prediction model needs to be incremental learning based on the preset incremental learning score threshold.

[0062] The implementation process of the HPC job power consumption prediction method based on online incremental learning of the present embodiment will be described in detail below.

[0063] The present embodiment adopts an online incremental learning job power consumption prediction framework, and the framework diagram is as shown in Figure 2 , the specific steps are as follows:

[0064] 1. Based on the offline HPC job power consumption data, the data sample is constructed using the sliding window, and the prediction model is trained offline through the constructed data sample to obtain the trained prediction model.

[0065] The prediction model is constructed based on a neural network, takes data stream as input, and outputs prediction of power consumption value at future time.

[0066] The sliding window constructs data samples by dividing HPC job power consumption data into multiple groups of data, and all historical data are used to initially train the prediction model.

[0067] 2. Based on the trained prediction model, the model parameters are updated to predict the new data at time t, and the predicted value is obtained.

[0068] 3. Calculate the error between the predicted value and the true value, and compare it with the threshold value. When the threshold value is exceeded, the incremental sample at time t is used to iteratively train the prediction model. Each iteration is based on the prediction model after the previous iteration, and incremental learning is performed to update the model parameters. If the threshold value is not exceeded, go directly to step 4.

[0069] 4. t = t + 1, go to step 2, and the loop condition ends when the end condition is met.

[0070] As the sampling work is carried out, the sample data will be updated continuously. Each training uses a combination of new and old samples, and the final model prediction capability is greatly improved in accuracy compared to traditional offline models.

[0071] For the timing judgment of incremental learning in step 3 above, the embodiment provides a judgment method based on model prediction performance decline score to perform data screening and realize selective updating of model parameters, thereby reducing the number of model parameter updates and improving the efficiency of model learning.

[0072] During the research on the online incremental learning model framework, it is found that there is a certain difference between the new sample data. The incremental data can have great similarity with the old data, but there is also a possibility of less similarity. Therefore, a necessary screening mechanism can realize selective updating of model parameters, reduce the number of model parameter updates, and improve the efficiency of model learning.

[0073] In the power consumption prediction simulation experiment, a model is usually used to predict future values, and the prediction results are compared with the actual values to evaluate the prediction performance of the model; when using online incremental learning to update the model, a suitable screening benchmark needs to be determined to select which data to use to update the model; a common screening benchmark is to set the prediction deviation value, which is the difference between the predicted value and the actual value, which can provide information about the accuracy and error size of the model; by setting an appropriate deviation value threshold, it can be determined which data can be applied to the update of the model to help the model better fit the data; but setting an appropriate threshold is a challenging task, when the threshold is set too low, some noise data will be included, which will cause the model to overfit, and when the threshold is set too high, some important data may be missed, which will cause the model to underfit.

[0074] The embodiment studies the update timing and conditions of the model, monitors and compares the "deviation distance" between the actual value and the predicted value in real time, and determines whether the prediction model needs to be updated based on the model update timing determination method. The average value of the sample error is not the best choice as the evaluation condition, and the embodiment determines whether the prediction model needs to be updated by monitoring the distribution change of the "deviation distance".

[0075] The embodiment uses a sliding window construction method to convert the "deviation distance" into a unified judgment standard, i.e. "model prediction performance decline score", the score is a number between [0, 1], the larger the score, the more serious the model performance decline, and vice versa. The better the prediction effect of the model, the basis for designing such a method is as follows:

[0076] (1) The amplitude dimension of the real world pattern data and the observed value is quite different, there are power consumption values of 70-150kw, and power consumption values of 140-480kw; therefore, the amplitude dimension of the deviation distance is also quite different, if the amplitude dimension of the observed data of the work power consumption value and the deviation dimension are unified into one measurement system, it will help to build a universal prediction model effect evaluation system.

[0077] (2) Although traditional absolute median deviation, 3Sigma / box plot, extreme value theory, etc. can be used to qualitatively determine that the prediction model needs to be updated when the deviation distance is higher than the threshold, but this method depends on the setting of the threshold, and different thresholds need to be set according to the amplitude dimension of different observed data; in addition, the real-time change of the data also requires the threshold to be updated, which increases the difficulty of application.

[0078] (3) The prediction model performance decline score is an interval value in [0%, 100%], which is more suitable for statistical modeling method to calculate the probability value from the statistical distribution of the "deviation distance". The probability is an interval value in [0, 1], which can be naturally used as the performance decline score of the prediction model.

[0079] Based on the above considerations, the specific calculation method of the model prediction performance decline score is as follows:

[0080] As shown in Figure 3 , two sliding data windows with lengths of large sliding window W and small sliding window W' are used, W is much larger than W', the deviation distance within the large sliding window W starting at t-k time and the deviation distance within the small sliding window W' starting at t time are buffered respectively, and the performance decline score of the prediction model at t time is obtained from the complementary probability of the right tail function (Q function) of the standard Gaussian distribution, as shown in Figure 4 , the specific formula of the model prediction performance decline score is as follows:

[0081]

[0082] Wherein, μ t and are the mean and variance of the "deviation distance" within the large sliding window W, is the mean of the "deviation distance" within the small sliding window W'.

[0083] The above formula can reflect the deviation of the data within the small sliding window W' relative to the data distribution within the large sliding window W. The larger the deviation is, the larger the anomaly score L t is, that is, the more mature the condition for updating the prediction model is; by setting the large and small sliding windows, the influence of short-term factors can be reduced, and the credibility of the model updating mechanism can be improved.

[0084] After the model prediction performance decline score is calculated, it is also necessary to determine whether the prediction model needs to be updated based on the preset incremental learning score threshold. Therefore, the setting of the incremental learning score threshold is related to the number of model updates, that is, the purpose of setting the threshold is to find a threshold, and then reduce the number of model updates while ensuring the prediction effect, so as to reduce the running time of the entire program. In the embodiment, two evaluation indexes are used to determine the incremental learning score threshold.

[0085] In the experiment, a threshold is set to determine when it is most appropriate to update the online incremental learning model parameters. Specifically, when the error between the predicted value and the actual value exceeds the set threshold, the prediction model parameters are updated using new data; the prediction model decline score is divided into 10 equal parts, and the best update threshold is found by using the two evaluation indexes of MAE and MAPE.

[0086] The MAPE result graph is as shown in Figure 5As shown in the MAPE result chart, when the threshold is set to 0.4, the running time of the program can be shortened to the greatest extent under the premise of ensuring the prediction accuracy; in the subsequent experiment, the threshold is set to 0.4, and the running time and accuracy of the online incremental learning model without threshold setting and the online incremental learning model with threshold setting are compared under the same conditions; the research results show that the prediction accuracy of the model with threshold setting and the model without threshold setting is basically the same, and the results are shown in Table 1, but the running time of the model with threshold setting is reduced by about 30% than the model without threshold setting, and the results are shown in Figure 6 Therefore, the research shows that the threshold setting is of great significance for developing more effective and fast online incremental learning model.

[0087] Table 1 MAPE and MAE values of the same job power consumption prediction with and without threshold setting

[0088]

[0089] Experimental verification

[0090] The prediction model of online incremental learning is verified by experiment. When the model is used to predict the power consumption of HPC jobs, new data is added to update the model parameters every time new data is received. Specifically, a sliding window strategy is adopted. In each update, the window is moved, and the new data is added to the window. The model parameters are updated by the new data. In this way, new data can be continuously learned without retraining the entire model parameters. In addition, the threshold is set to control the update mechanism of the online incremental learning model. When the error between the new data and the prediction result is greater than the threshold, the model parameters are updated. While ensuring the accuracy of the model, new data can also be timely incorporated into the model training, realizing the rapid adaptation of data. This ensures efficient performance while effectively avoiding model performance degradation. Based on the above settings, the present embodiment carries out offline training and online training, Figure 7 is a flowchart of offline training, Figure 8 is a flowchart of online training.

[0091] Experimental results and analysis

[0092] In order to test the performance of the HPC job power consumption framework based on online incremental learning, the power consumption data of an existing HPC cluster is used for experiment. The experiment is divided into two cases: one is to predict the same type of job, which is manually classified according to the existing job power consumption, aiming to verify the performance of the framework in the case of high similarity; the other is to mix different jobs for prediction, because in practical application, HPC platform jobs cannot be classified in detail, and the number of available jobs is limited.

[0093] The embodiment first carries out data preprocessing, and verifies the experiment by using "leave-one-out method"; the experiment adopts four models, including LSTM, GRU, Seq2seq and NBeats; the data is verified in different types, including Case1 to Case3, which have the same points of small data amount and high job similarity, while Case4 to Case5 are used for prediction under the condition of large data amount; the data set of Case4 is similar job, and the data set of Case5 is mixed job.

[0094] The online incremental learning model and the original model are compared, and the experimental results show that, whether the same type of job or the different type of job, the model using online incremental learning has certain improvement than the original model, especially the GRU and NBeats models have the largest improvement range; the MAE and MAPE values are used to show the experimental results, and table 2 and table 3 show the MAE and MAPE results respectively.

[0095] Table 2 MAE values of online incremental update prediction and offline prediction of different models

[0096]

[0097]

[0098] Table 3 MAPE values of online incremental update prediction and offline prediction of different models

[0099]

[0100] Embodiment two

[0101] In one or more embodiments, an HPC job power consumption prediction system based on online incremental learning is disclosed, including an initialization module, a data acquisition module, a power consumption prediction module and an online learning module:

[0102] The initialization module is configured to initialize the prediction model according to the offline HPC job power consumption data;

[0103] The data acquisition module is configured to acquire the newly sampled HPC job power consumption data in real time and form an input data stream;

[0104] The power consumption prediction module is configured to process the input data stream by the prediction model to generate and output the power consumption prediction value at the future time;

[0105] The online learning module is configured to update the model by using online incremental learning based on the model update timing determination method during processing the input data stream;

[0106] The model updating occasion judgment method is based on a preset incremental learning score threshold to determine whether the prediction model needs to be subjected to incremental learning by monitoring the deviation distance distribution change between the true value and the predicted value and obtaining a model prediction performance decline score when new data is received.

[0107] Embodiment three

[0108] An object of the present embodiment is to provide a computer-readable storage medium.

[0109] A computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps in the HPC job power consumption prediction method based on online incremental learning as described in Embodiment One of the present disclosure.

[0110] Embodiment four

[0111] An object of the present embodiment is to provide an electronic device.

[0112] An electronic device includes a memory, a processor, and a program stored on the memory and executable on the processor, and the processor implements the steps in the HPC job power consumption prediction method based on online incremental learning as described in Embodiment One of the present disclosure when executing the program.

[0113] The above merely provides preferred embodiments of the present application but not for limiting the present application. For those skilled in the art, the present application can have various modifications and changes. Any modified, equivalent replaced, improved, and the like within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting HPC job power consumption based on online incremental learning, characterized in that, include: Initialize the prediction model based on offline HPC job power consumption data; Real-time acquisition of newly sampled HPC job power consumption data and formation of an input data stream; The prediction model processes the input data stream to generate and output power consumption predictions for future times. During the processing of the input data stream, online incremental learning is used to update the model based on the model update timing determination method. The method for determining the timing of model updates involves monitoring the change in the deviation distance distribution between the true and predicted values ​​when new data is received to obtain a score indicating a decline in model prediction performance. Based on a preset incremental learning score threshold, it determines whether the prediction model needs to perform incremental learning. The model prediction performance degradation score is constructed using a sliding window method, which converts the deviation between the predicted value and the true value into a unified judgment standard in real time, namely the model prediction performance degradation score. The larger the score, the more severe the degradation of the model performance, and vice versa. The method of constructing using sliding windows is as follows: Two sliding data windows of varying lengths are used, one large and one small, to cache the deviation distances before and after time step, respectively. The performance degradation score of the time-based prediction model is obtained from the complementary probabilities of the right-tail function of the standard Gaussian distribution. The specific formula for the performance degradation score of the model prediction is as follows: in, and For large sliding windows The mean and variance of the deviation distance within the range, For small sliding windows The mean deviation distance within.

2. The HPC job power consumption prediction method based on online incremental learning as described in claim 1, characterized in that, The initialization prediction model is specifically as follows: Based on offline HPC job power consumption data, a sliding window is used to construct data samples. The constructed data samples are then used to train the prediction model offline, resulting in a well-trained prediction model.

3. The HPC job power consumption prediction method based on online incremental learning as described in claim 1, characterized in that, The prediction model is built on a neural network, takes a data stream as input, and outputs a prediction of power consumption at a future time.

4. The HPC job power consumption prediction method based on online incremental learning as described in claim 1, characterized in that, The specific processing steps upon receiving new data are as follows: New data is predicted using a predictive model to obtain predicted values. Calculate the model's predictive performance degradation score based on the distance between the predicted and actual values; The model prediction performance degradation score is compared with the incremental learning score threshold. When the incremental learning score threshold is exceeded, new data is used as incremental samples for learning, and the parameters of the prediction model are updated.

5. The HPC job power consumption prediction method based on online incremental learning as described in claim 1, characterized in that, The incremental learning score threshold is set by dividing the model prediction performance degradation score into multiple equal parts, and using the two evaluation metrics MAE and MAPE to find the optimal time for model update, i.e., the incremental learning score threshold.

6. A power consumption prediction system for HPC jobs based on online incremental learning, characterized in that, It includes an initialization module, a data acquisition module, a power consumption prediction module, and an online learning module: The initialization module is configured to initialize the prediction model based on offline HPC job power consumption data. The data acquisition module is configured to: acquire newly sampled HPC job power consumption data in real time and form an input data stream; The power consumption prediction module is configured such that the prediction model processes the input data stream to generate and output power consumption prediction values ​​for future times. The online learning module is configured to update the model using online incremental learning based on the model update timing determination method during the processing of the input data stream. The method for determining the timing of model updates involves monitoring the change in the deviation distance distribution between the true and predicted values ​​when new data is received to obtain a score indicating a decline in model prediction performance. Based on a preset incremental learning score threshold, it determines whether the prediction model needs to perform incremental learning. The model prediction performance degradation score is constructed using a sliding window method, which converts the deviation between the predicted value and the true value into a unified judgment standard in real time, namely the model prediction performance degradation score. The larger the score, the more severe the degradation of the model performance, and vice versa. The method of constructing using sliding windows is as follows: Two sliding data windows of varying lengths are used, one large and one small, to cache the deviation distances before and after time step, respectively. The performance degradation score of the time-based prediction model is obtained from the complementary probabilities of the right-tail function of the standard Gaussian distribution. The specific formula for the performance degradation score of the model prediction is as follows: in, and For large sliding windows The mean and variance of the deviation distance within the range, For small sliding windows The mean deviation distance within.

7. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-5.

8. A storage medium, characterized in that, The computer-readable instructions are stored non-transitory, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-5 are executed.

Images