A data center computing power load demand combination prediction method and system considering PUE

Abstract

The present invention discloses a combined prediction method and system for data center computing power load demand based on a neural network and considering PUE. The system includes a data center computing power load demand prediction model construction module, a BP neural network optimization module, a model training module, and a prediction module. The present invention uses a BP-GRU combined prediction model to improve the traditional prediction method, and adopts an improved GA-PSO algorithm to improve the defects of the BP neural network such as sensitivity to initial weights, easy to fall into local optimality, slow convergence speed, and poor generalization ability. It also combines the GRU model for combined prediction, making up for the lack of temporal state defects of the BP neural network. Based on the demand side perspective, this prediction method considers the linkage between energy consumption indicators in the data center and various internal and external influencing factors, providing a feasible solution for the prediction of computing power load demand of data center enterprises, in order to carry out scientific and effective cross-domain scheduling.

Description

Technical Field

[0001] The present invention belongs to the field of computer science information analysis and prediction technology, and specifically relates to a data center computing power load demand combination prediction method and system considering PUE. Background Art

[0002] The digital economy is becoming a key force in restructuring resource factors, reshaping economic structures, and enhancing market competitiveness. Data centers transform electricity into computing power, driving the development of the digital economy. Given the challenges of achieving efficient, large-scale storage and the ever-changing load demands of power storage technologies, accurate forecasting of data center computing load demand is an effective method for achieving a dynamic balance between power system supply and computing load demand. This helps data center companies increase their availability, reduce resource waste costs, optimize energy consumption indicators, and facilitate scheduling planning.

[0003] Currently, data center computing load demand primarily impacts the IT equipment used to store and transmit data. Due to its high variability and heterogeneity, this load is influenced by multiple factors. Data center load forecasting studies typically analyze the total load from a supply-side perspective, selecting indicators such as temperature, cooling capacity, heating capacity, humidity, and internal and external data center equipment for a single mapping analysis. This approach neglects the impact of data center energy consumption indicators and the demand side, and fails to consider the interconnectedness of these factors. This results in an incomplete analysis of the characteristics of computing load demand.

[0004] Furthermore, traditional forecasting models, such as time series models, neural network models, and artificial intelligence optimization models, offer good prediction results on a single dataset, but are less accurate when used across multiple datasets. Furthermore, traditional forecasting algorithms are prone to defects such as sensitivity to initial weights, local optimal solutions, slow convergence, and poor generalization. These issues pose challenges to accurate load forecasting in data centers.

[0005] In the existing patent, "A Multi-Load Forecasting Method for Data Center Stations Based on Hybrid Model Prediction (Application No. CN202111048836.5)" uses the GRA method to perform feature analysis and normalization on multi-load data, and inputs the processed data into the QPSO-BP neural network and XGBoost model for parallel prediction. However, it only considers external factors affecting the power load of the data center, and does not combine key influencing factors such as the PUE value of the internal energy consumption of the data center for prediction. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems of ignoring the impact of data center energy consumption indicators and demand side, and not considering the linkage between various influencing factors, and the traditional prediction algorithm is very likely to fall into defects such as abnormal sensitivity to initial weights, local optimal solutions, slow convergence speed, and poor generalization ability, the present invention proposes a combined prediction method for data center computing power load demand considering PUE. On the basis of considering the external weather factors of the data center, the influence of the PUE value of the internal energy consumption of the data center is also considered, and the K-Means algorithm is used for cluster analysis of the data set for feature extraction. Then, the BP-GRU combined prediction model is used to improve the traditional prediction method, and the GA-PSO algorithm is introduced to optimize the defects of the BP neural network, and machine learning and deep learning technologies are simultaneously applied to load forecasting.

[0007] A method for predicting data center computing power load demand based on PUE to achieve one of the objectives of the present invention includes the following steps:

[0008] S1. Establishing a data center computing load demand prediction model based on historical workload data generated by a large-scale GPU cluster within a preset time period. The data center computing load demand prediction model is a combined prediction model of a BP neural network and a GRU neural network optimized using a GA-PSO optimization algorithm, and is used to predict computing load demand at future moments based on current workload data.

[0009] The workload data includes date, user ID, number of task instances running in parallel corresponding to the date and the running time of each task instance, number and model of GPUs;

[0010] The input layer data of the data center computing load demand prediction model includes: computing load demand, PUE value, and time characteristic value; the output layer data is computing load demand;

[0011] The time characteristic value is used to determine whether the date in the workload data is a holiday or a working day; holidays and working days are marked with different tags respectively;

[0012] The calculation formula of the PUE value is:

[0013] PUE = total energy consumption of all machines in the data center / total energy consumption of GPUs in the data center;

[0014] The computing load requirement is the number of task instances running in parallel in the data center;

[0015] S2. Optimize the weights and thresholds of each dimension of the BP neural network using the GA-PSO optimization algorithm to obtain a GA-PSO-BP network model. The optimization objective is the inverse of the sum of squared sample errors, so as to minimize the fitting error of the data center computing power load demand prediction model.

[0016] The sample is the historical data of the workload generated by the GPU cluster; when the inverse of the sum of squared errors of the samples no longer decreases, it is considered that the weights and thresholds of each dimension of the BP neural network are optimal at this time, and the GA-PSO-BP network model training is completed;

[0017] S3. Divide the historical data of the workload generated by the large-scale GPU cluster into a training set and a test set, use the training set to train the data center computing power load demand prediction model to obtain a trained data center computing power load demand prediction model, and in the testing phase, input the test set into the GA-PSO-BP network model; after reaching the maximum number of iterations, output the predicted computing power load demand A; then input the computing power load demand A into the GRU neural network to obtain a predicted computing power load demand data set B; after reaching the maximum number of iterations, multiply the data set A with the weight of the output result of the GA-PSO-BP network model, and multiply the data set B with the weight of the output result of the GRU neural network, and then use the sum of the products as the predicted computing power load demand output, evaluate the predicted computing power load demand output, and determine whether the trained data center computing power load demand prediction model meets the requirements.

[0018] S4. Input the current workload data into the data center computing power load demand prediction model trained in step S3. The data center computing power load demand prediction model outputs the computing power load demand of the data center at future times.

[0019] Furthermore, in order to improve the deficiency of BP neural network in lacking time series state, GRU neural network is combined with BP neural network to perform combined prediction of neural network.

[0020] Furthermore, in step S2, the GA-PSO algorithm is used to optimize the weights and thresholds of the BP neural network, and the following parameters need to be set:

[0021] (1) Set the computing load demand, PUE value, and time characteristics in the workload data as the input layer of the BP network; set the computing load demand as the output layer of the BP neural network;

[0022] (2) Set the number of hidden layers of the BP network according to the following formula:

[0023]

[0024] Where:

[0025] H is the number of hidden layer nodes;

[0026] M is the number of nodes in the input layer; N is the number of nodes in the output layer;

[0027] a is a set adjustment integer, the preferred range is [1,10];

[0028] (3) Set the learning rate η of the BP network according to the following formula:

[0029]

[0030] Where:

[0031] T now is the current iteration number;

[0032] T max is the maximum number of iterations;

[0033] (4) Set the tansig function as the node transfer function between the output layer and the hidden layer; the purelin linear function as the node transfer function of the output layer; and the trainlm function as the training function;

[0034] The crossover and mutation operations in the GA algorithm are used to optimize the parameters of the GA-PSO algorithm. The specific optimization process is as follows:

[0035] The velocity v of particle i in dimension d (1≤d≤D) id and position X id As the individual extreme value and the group extreme value are updated, the update formulas are as follows:

[0036]

[0037] X id =X id +v id

[0038] Where:

[0039] V id : represents the velocity of the i-th particle in the d-th dimension;

[0040] X id : represents the position of the i-th particle in the d-th dimension;

[0041] ω is the inertia weight;

[0042] c1, c2 are learning factors;

[0043] Represents a random number in the interval [0,1], where t is the current iteration number;

[0044] Pbest id is the individual extreme value of the i-th particle in the d-th dimension, where d is the spatial dimension of the solution;

[0045] Gbestid is the group extreme value of the i-th particle in the d-th dimension, where d is the spatial dimension of the solution;

[0046] During the iteration process, the quality of each particle needs to be evaluated. A linear function of the inverse of the sum of squared sample errors is used as the fitness function. To prevent accidental mutation failure or crossover failure, the present invention introduces an inertia weight to optimize the fitness function. The fitness calculation formula is:

[0047]

[0048] Where:

[0049] fitness now : represents the updated fitness value;

[0050] fitness past : Indicates the fitness value before updating;

[0051] w now +w past =1, A and B are both set constants;

[0052] p i is the network prediction value of the i-th training sample;

[0053] r i is the actual output value of the i-th training sample;

[0054] N is the total number of training samples.

[0055] Particle swarm crossover and mutation operations: Real number encoding is performed on the individuals in the population, with the crossover probability P c Make a cross.

[0056] Set the number of crossover segments to n, where n is an integer between 0 and 10; divide the chromosome into 2n segments, and crossover the odd-numbered segments; that is, if the number of crossover segments is 4, then divide the chromosome into 8 segments, of which segments 1, 3, 5, and 7 are crossed.

[0057] The crossover operation for paired particles i and j is as follows:

[0058] Speed ​​Cross:

[0059] Position Crossover:

[0060] Where:

[0061] α1, α2 are random numbers in the interval [0,1];

[0062] V jd : represents the velocity of the jth particle in the dth dimension;

[0063] X jd : represents the position of the jth particle in the dth dimension;

[0064] Select the i-th particle with mutation probability P m Perform mutation as follows:

[0065] Speed ​​variation:

[0066] Position variation:

[0067] Where:

[0068] V id (t): represents the velocity of the i-th particle at time t in the d-th dimension;

[0069] X id (t): represents the position of the i-th particle at time t in the d-th dimension;

[0070] v min , v max are the minimum and maximum velocities of the particle, respectively;

[0071] X min , X max are the lower and upper bounds of the particle position, respectively;

[0072] is a random number in the interval [0,1].

[0073] A data center computing power load demand combined prediction system considering PUE to achieve the second purpose of the present invention includes a data center computing power load demand prediction model construction module, a BP neural network optimization module, a model training module and a prediction module;

[0074] The data center computing power load demand prediction model construction module is used to establish a data center computing power load demand prediction model based on the historical data of the workload generated by the GPU cluster within a preset time period. The data center computing power load demand prediction model is a combined prediction model of the BP neural network and the GRU neural network optimized based on the GA-PSO optimization algorithm, and is used to predict the computing power load demand at future moments based on the workload data at the current moment;

[0075] The BP neural network optimization module is used to optimize the weights and thresholds of each dimension of the BP neural network using the GA-PSO optimization algorithm to obtain a GA-PSO-BP network model. Its optimization objective is the inverse of the sum of squared sample errors, so as to minimize the fitting error of the data center computing power load demand prediction model.

[0076] The model training module is used to divide the historical data of the workload generated by the GPU cluster into a training set and a test set, and use the training set to train the data center computing power load demand prediction model to obtain a trained data center computing power load demand prediction model. In the testing phase, the test set is input into the GA-PSO-BP network model; after reaching the maximum number of iterations, the predicted computing power load demand A is output; the computing power load demand A is then input into the GRU neural network to obtain a predicted computing power load demand data set B; after reaching the maximum number of iterations, the data set A is multiplied by the weight of the output result of the GA-PSO-BP layer to obtain a product A', and the data set B is multiplied by the weight of the output result of the GRU layer to obtain a product B', and the sum of A' and B' is used as the predicted computing power load demand output, and the predicted computing power load demand output is evaluated to determine whether the trained data center computing power load demand prediction model meets the requirements;

[0077] The prediction module inputs the current workload data into the trained data center computing power load demand prediction model, and the data center computing power load demand prediction model outputs the computing power load demand of the data center at future times.

[0078] Furthermore, the BP neural network optimization module also includes a parameter optimization module and a particle optimization module;

[0079] The parameter optimization module is used to optimize the number of hidden layers of the BP network, the learning rate of the BP network, the node transfer function between the output layer and the hidden layer, the node transfer function of the output layer, and the training function;

[0080] The particle optimization module is used to calculate the velocity v of particle i in d dimension (1≤d≤D) in BP neural network. id and position X id Optimize.

[0081] The BP neural network optimization module also includes a particle crossover optimization module for optimizing the speed and position of the paired particles i and j in the crossover process of the BP neural network by using the crossover operation in the GA algorithm.

[0082] The BP neural network optimization module also includes a particle mutation optimization module for optimizing the speed and position of particles in the mutation process of the BP neural network by using the mutation operation in the GA algorithm.

[0083] Beneficial effects

[0084] This paper improves on traditional prediction methods by using a BP-GRU combined prediction model. It also employs an improved GA-PSO algorithm to address the shortcomings of the BP neural network, such as its sensitivity to initial weights, tendency to fall into local optimality, slow convergence, and poor generalization. This algorithm, combined with the GRU model for combined prediction, overcomes the BP neural network's lack of temporal state. This prediction method, based on the demand side, considers the linkage between energy consumption indicators within the data center and various internal and external influencing factors. It provides a feasible solution for predicting computing power load demand for data center enterprises in the context of "Eastern Data and Western Computing," enabling scientific and effective cross-domain scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is a schematic flow chart of an embodiment of the method of the present invention;

[0086] Figure 2 It is a schematic diagram of the BP neural network structure;

[0087] Figure 3 1 is a flow chart of optimizing a BP neural network model using a GA-PSO optimization algorithm in an embodiment of the method of the present invention;

[0088] Figure 4 This is a schematic diagram of the GRU structure. DETAILED DESCRIPTION

[0089] The following detailed description is intended to explain the technical solutions of the present invention claims, so that those skilled in the art can understand the present claims. The scope of protection of the present invention is not limited to the specific implementation structures described below. Any implementation schemes created by those skilled in the art that incorporate the technical solutions of the present invention claims but differ from the following detailed descriptions are also within the scope of protection of the present invention.

[0090] The following combination Figures 1 to 4 An embodiment of the method of the present invention is described below.

[0091] Step 1: Collect the original data set generated by the large-scale GPU cluster in the data center. The data set includes the date, user ID, number of task instances and their runtime, number and model of GPUs. Convert the monthly data set into hourly units. Obtain the corresponding PUE (Power Usage Effectiveness, an indicator for evaluating the energy efficiency of data centers). The computing load demand in this invention is the user's task instance on the platform. The PUE value is the total energy consumption of all machines in the data center divided by the total energy consumption of the GPUs in the data center. Relevant research shows that PUE is affected by temperature; the higher the temperature, the greater the PUE value.

[0092] Step 2: Preprocess the original data set. Given the large amount of diverse user data, each data set needs to be cleaned separately, including: expanding the feature attributes of the data set, processing missing values ​​and outliers, and normalization.

[0093] Dataset feature expansion means incorporating feature attributes that may affect the computing load requirements of the data center into the current user's dataset.

[0094] For example, for weekdays and holidays on a date feature, if it is a holiday, an additional value of 1 is assigned, and if it is a weekday, an additional value of 0 is assigned. This additional value is used as an extended attribute of the time feature dataset. The extended attribute is merged with the original dataset to form a dataset with extended features.

[0095] The handling of missing values ​​and outliers includes: filling missing values ​​with fixed values ​​and removing outliers such as negative values;

[0096] A data center computing load demand prediction model is established based on the historical data of workloads generated by large-scale GPU clusters within a preset time period. The input nodes of the prediction model are the data center computing load demand value, PUE value, and time characteristics of the day before the prediction date in hours. The output node is the data center computing load demand value of the predicted day. Given that the PUE value range is generally between [1, 2] and has a large difference from the data center computing load demand value, the feature scaling method is used to normalize the samples. The feature value X of each data is set as i Scale to [0,1] or [-1,1], that is:

[0097]

[0098] Among them, X max , X min are the maximum and minimum values ​​of the eigenvalue X, respectively. i is the original data, X i is the i-th value of eigenvalue X.

[0099] Step 3: K-means clustering model optimization. The K-means clustering model is used to cluster similar computing load curves at different times to improve the accuracy of the subsequent data center computing load demand prediction model. Considering the impact of different users' personalized needs on computing load, this paper uses user ID as a feature, re-splitting the normalized dataset into multiple datasets based on user ID, and then clustering the split datasets using the K-means clustering model.

[0100] In response to the problem of selecting the k value in the K-Means clustering model, in this embodiment, a k-mark curve composed of different k values ​​and the loss function values ​​of the clustering model is drawn. As the k value increases, the mark will decrease, and the number of samples contained in each class will decrease, so the sample is closer to the center point of its class. As the k value continues to increase, although the mark will still decrease, the rate of decrease will obviously slow down, which means that the improvement effect of the clustering model will continue to decrease. In the process of increasing the k value, the value corresponding to the position where the improvement effect of the clustering model decreases the most is the turning point in the curve. In the present invention, this k value is used as the approximately optimal k value. Further, the mark calculation formula in the present invention is as follows:

[0101]

[0102] Among them, mark is the sum of squares of errors from each sample point to the center of gravity, m i Expressed as the center of gravity of the i-th class, P i is m i Sample points in the class.

[0103] Step 4: GA-PSO-BP neural network model optimization. The data set optimized by the K-Means clustering model is randomly divided into 7:3 ratios for training and testing of the neural network model. The GA-PSO optimization algorithm is used to optimize the weights and thresholds of the BP neural network model to obtain the GA-PSO-BP network model. Its optimization target is the inverse of the sum of squares of sample errors. The optimization process is as follows: Figure 3 As shown in the figure, a computing load demand prediction model with a small fitting error is obtained; the dimension D is the sum of the BP neural network weight and threshold. The specific process of GA-PSO optimizing the BP neural network is as follows:

[0104] 1. Initialize the BP neural network topology. Determine the number of neurons in the input layer, hidden layer, and output layer and the training function. Figure 2 This is the BP neural network structure diagram. The computing load demand, PUE value, and time characteristics (whether it is a holiday) are used as the input of the BP neural network, so the number of input layer neurons is 3; the computing load demand is used as the output of the BP neural network, so the number of output layer neurons is 1;

[0105] 2. The appropriate number of hidden layer nodes H is the key to the smooth implementation of the BP neural network structure function. The basic principle for determining the number of hidden layer nodes is: under the premise of meeting the accuracy, take the most compact structure possible, that is, take the least number of hidden layer nodes possible. This invention uses the rounding up method to determine the optimal number of hidden layers: Among them, H is the number of hidden layer nodes, M is the number of input layer nodes, N is the number of output layer nodes, and a is an adjustment integer between 1 and 10.

[0106] 3. The learning rate η is the amount of correction change for updating weights and thresholds in each training of the BP neural network. If the learning rate η is too small, it may lead to slow convergence or fall into a suboptimal solution. If the learning rate η is too large, the optimal solution may be missed due to the large update step size, and the network structure may fall into an unstable oscillation period. Considering that the range of the learning rate η is less than 1 and the value is relatively small, this embodiment makes a dynamic transformation calculation of the learning rate: Among them, T now is the current iteration, T max is the maximum number of iterations.

[0107] 4. To address nonlinearity, the BP neural network's input and hidden layers often use nonlinear functions, such as the S-shaped transfer function, to keep intermediate process data within a reasonable range. The output layer uses a linear function, allowing the output data to take on any value. In this embodiment, the nonlinear tansig function is selected as the node transfer function between the output and hidden layers, while the purelin linear function is used as the node transfer function for the output layer. The trainlm function is used as the training function.

[0108] 5. Design of BP neural network weights and thresholds. This embodiment uses error correction learning rules to train the BP neural network, as shown below:

[0109] 5.1. Forward calculation process of the signal: the calculation process in which sample data propagates from the input layer through the hidden layer to the output layer, and the network output value is calculated based on the input sample.

[0110] The input Hid of the i-th node in the hidden layer of the network i :

[0111]

[0112] The output y of the i-th node in the hidden layer of the network i :

[0113]

[0114] The input of the kth node in the network output layer is In k :

[0115]

[0116] The output of the kth node in the network output layer is Out k :

[0117]

[0118] In the above formula:

[0119] Hidi is the input of the i-th node in the hidden layer of the network;

[0120] w ij is the connection weight between hidden layer node i and input layer node j;

[0121] x j is the input of the jth neuron in the input layer;

[0122] M is the number of neurons in the hidden layer;

[0123] b i is the threshold of hidden layer node i;

[0124] y i is the output of the i-th node in the hidden layer of the network;

[0125] f1 is the transfer function of the hidden layer; f2 is the transfer function of the output layer;

[0126] In k is the input of the kth node in the network output layer;

[0127] a k is the threshold of the output layer;

[0128] Out k is the output of the kth node in the network output layer;

[0129] S is the number of neurons in the output layer.

[0130] 5.2. Error signal reverse transmission calculation process:

[0131] The network output error is obtained by using the sample forward transfer calculation process to learn and correct the network weights and thresholds until the expected value or the number of iterations reaches the output error. The error function is:

[0132]

[0133] In the above formula:

[0134] E is the sample error;

[0135] w is the inertia weight;

[0136] b is the dimension;

[0137] T k Iterate the target output values ​​of k nodes for the output layer;

[0138] Out k Iterate over the actual output values ​​of k nodes for the output layer.

[0139] According to the gradient descent method, the weight of the corrected output layer is calculated in reverse. The weight from the i-th input to the k-th output is:

[0140]

[0141] The hidden layer weights are modified, and the weights from the j-th input to the i-th output are:

[0142]

[0143] Output layer threshold correction:

[0144]

[0145] Hidden layer threshold correction:

[0146]

[0147] according to:

[0148]

[0149] The final weight and threshold adjustment formula is:

[0150]

[0151]

[0152]

[0153]

[0154] In the above formula:

[0155] Δw ki is the weight from the i-th input to the k-th output;

[0156] η is: learning rate;

[0157] s is: the number of neurons in the output layer;

[0158] T k Iterate the target output value of k nodes for the output layer;

[0159] OUT k is the output of the kth node in the network output layer;

[0160] f'1(Hid k ) is the inverse of the transfer function of the hidden layer K nodes;

[0161] y i is the output of the i-th node in the hidden layer of the network;

[0162] f2'(Hid i ) is the inverse of the transfer function of the output layer K nodes;

[0163] x j is: the input of the jth neuron in the input layer;

[0164] Δw ij is: the weight from the jth input to the i-th output;

[0165] Δa k is: the corrected output layer threshold;

[0166] Δb i is: the modified hidden layer threshold;

[0167] The network output deviation is passed from the output layer to the hidden layer and then to the input layer, and the weights and thresholds of each layer of the network are corrected.

[0168] 6. Population initialization. To improve the defect of BP neural network being sensitive to weights, all parameters of BP neural network are regarded as a D-dimensional vector, which makes it easier to introduce GA-PSO algorithm to optimize the weights and thresholds of each dimension of BP network. The basic parameters of GA-PSO algorithm are set as follows:

[0169] ① The number of particles in population U is set to m; population U represents the data sequence of the input factors affecting the computing load demand. In this example, the population number U is set to 30 and the number of particles m is set to 100. After the particles are encoded with real numbers, they correspond to the weights and thresholds of the BP neural network. That is, one encoded information contains all the weights and thresholds of a BP neural network and represents each individual in the population.

[0170] ③ Upper and lower limits of particle speed v min ,~v max :The vmin,~vmax represent the range width of the particle moving distance; in this example, v min Set to -5, v max Set to 5;

[0171] ④ Set the learning factors c1 and c2: the preferred range is [0, 4], and the preferred value is 2;

[0172] ⑤ Set the inertia weight ω: the preferred range is [0.4, 0.9], and the preferred value is 0.8;

[0173] ⑥ Set the convergence accuracy to δ, depending on the specific problem;

[0174] ⑦ Set the maximum number of iterations to G. Adjust it according to the actual problem. In this example, it is set to 100.

[0175] ⑧Set the crossover probability Pc , mutation probability P m : When the particle swarm is iterated, the position vector and velocity vector of the particle are cross-operated and mutated with a certain probability to avoid falling into the local extreme point, thereby enhancing the spatial search performance of the population and the diversity of the population. In this embodiment, P c The preferred value of P is 0.3. m The preferred value of is 0.1, but is not limited thereto.

[0176] 7. Use the crossover and mutation operations in the GA algorithm to optimize the parameters of the GA-PSO algorithm. The specific optimization process is as follows:

[0177] The velocity V of particle i in dimension d (1≤d≤D) id and position X id It changes with the update of individual extreme values ​​and group extreme values, where D is the total dimensional data, and the update formulas are as follows:

[0178]

[0179] X id =X id +v id

[0180] Where t is the current iteration number; d is the spatial dimension of the solution; ω is the inertia weight; C1 and C2 are learning factors; Represents a random number in the interval [0,1]; Pbest id is the dth dimension of the individual extreme value of the i-th particle; Gbest id is the d-th dimension of the group extreme value.

[0181] Fitness function design: During the iteration process, the quality of each particle needs to be evaluated. A linear function of the inverse of the sum of squared sample errors is used as the fitness function. To prevent accidental mutation failure or crossover failure, the present invention introduces an inertia weight, which is calculated as follows:

[0182]

[0183] Where:

[0184] fitness now : represents the updated fitness value;

[0185] ω now : represents the updated inertia weight;

[0186] fitness past : Indicates the fitness value before updating;

[0187] ω past and ωnow : Represents the inertia weight before and after iteration, w now +w past =1;

[0188] A and B are both set constants;

[0189] P i is the network prediction value of the i-th training sample;

[0190] r i is the corresponding actual output value;

[0191] N is the total number of training samples.

[0192] Particle swarm crossover and mutation operations: Real number encoding is performed on the individuals in the population, with the crossover probability P c Crossover, in this embodiment, P c It is 0.3, but not limited to this.

[0193] Set the number of crossover segments to n. In this embodiment, n is an integer between 0 and 10. The number of crossover segments is selected by the rand function. After the number of crossover segments is selected, the chromosome is divided into 2n segments, and the odd segments are crossed. That is, if the number of crossover segments is 4, the chromosome is divided into 8 segments, of which segments 1, 3, 5, and 7 are crossed. The crossover operation for paired particles i and j is as follows:

[0194] Speed ​​Cross:

[0195] Position Crossover:

[0196] Where:

[0197] α1 and α2 are random numbers in the interval [0,1];

[0198] V jd : represents the velocity of the jth particle in the dth dimension;

[0199] X jd : represents the position of the jth particle in the dth dimension;

[0200] Select the i-th particle for mutation as follows:

[0201] Speed ​​variation:

[0202] Position variation:

[0203] Where, v min 、v max are the minimum and maximum speeds of the particles respectively; X min 、X maxare the lower and upper bounds of the particle position, respectively; is a random number in the interval [0,1].

[0204] Step 5: Consider the combined prediction model for data center computing power load demand based on PUE. This combined prediction model is based on a combination of the GA-PSO-BP model and the GRU model. The Gated Recurrent Unit (GRU) is a type of recurrent neural network (RNN) and a simplified and improved version of the Long Short-Term Memory (LSTM) network. The GRU model introduces two gate functions: an update gate and a reset gate, which are used to memorize and filter information, respectively. The GRU mathematical model is as follows:

[0205] z t =σ(U (z) h t-1 +W (z) x t )

[0206] r t =σ(U (r) h t-1 +W (r) x t )

[0207]

[0208] h t =(1σ-z t )⊙h t-1 +z t ⊙h t

[0209] y t =σ(W'h t )

[0210] Where σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, and U (z) and W (z) is the weight matrix of the update gate, U (r) and W (r) is the weight matrix of the reset gate, h t is the output of the current layer at time t, x t is the PUE value and time characteristics of the input of the hidden layer at time t, y t The predicted computing load demand. is the previous moment h t-1 Output and current input x t The specific internal structure is as follows. Figure 4 shown.

[0211] Set the GRU layer parameters as follows:

[0212] ① The sliding time window length is set to 24, that is, the length of the time segment used for prediction is equal to the output dimension of the GA-PSO-BP network, that is, the sequence output of the data center computing power load for 24 consecutive hours is used as an output dimension;

[0213] ② The training batch size is set to 56, which is the number of samples selected before each parameter adjustment;

[0214] ③The GRU input dimension is set to be the same as the output dimension of the GAP-SO-BP network;

[0215] ④The number of neurons in the GRU layer is set to 60;

[0216] ⑤ The weight of the GA-PSO-BP layer output result and the weight of the GRU layer output result are 0.4 and 0.6 respectively; the weight of the GA-PSO-BP layer output result is the GRU layer input weight;

[0217] ⑥The number of iterations of the forecast combination model is set to 100;

[0218] The historical workload data generated by the large-scale GPU cluster is divided into a training set and a test set in a ratio of 7:3. First, the preprocessed training set is input into the GA-PSO-BP network model and the GRU neural network model for training, and the trained model parameters are saved. The training results are used to predict the test set. During the testing phase, the test set is input into the GA-PSO-BP network. After reaching the maximum number of iterations, the predicted computing load demand A is output. The computing load demand A is then input into the GRU neural network to obtain the predicted computing load demand dataset B. After reaching the maximum number of iterations, the sum of datasets A and B multiplied by their own result weights is output as the predicted computing load demand.

[0219] Step 6: Model training and testing. The combined prediction model is trained using the training set of computing load requirements and its MSE (mean square error) is calculated. The model is tested using the test set of computing load requirements and the MSE of the model output is compared with that of the test sample. The calculation formula is: Where N is the total number of training samples, n=1...N,y i Indicates the actual computing load demand, Indicates the predicted computing load demand.

[0220] Step 7, model evaluation. In order to better demonstrate the superiority of the combined model, the present invention selects the mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE) as three performance index functions for error comparison analysis. The calculation formulas for MAE and RMSE are:

[0221] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0222] The embodiment of the present application also provides a data center computing power load demand combined prediction system considering PUE, including a data center computing power load demand prediction model construction module, a BP neural network optimization module, a model training module and a prediction module;

[0223] The data center computing load demand prediction model construction module is used to establish a data center computing load demand prediction model based on the historical workload data generated by the GPU cluster within a preset time period. The data center computing load demand prediction model is a combined prediction model of the BP neural network and the GRU neural network optimized based on the GA-PSO optimization algorithm, and is used to predict the computing load demand at future moments based on the workload data at the current moment;

[0224] The BP neural network optimization module is used to optimize the weights and thresholds of each dimension of the BP neural network using the GA-PSO optimization algorithm to obtain a GA-PSO-BP network model. Its optimization objective is the inverse of the sum of squared sample errors, so as to minimize the fitting error of the data center computing power load demand prediction model.

[0225] The model training module is used to divide the historical data of the workload generated by the GPU cluster into a training set and a test set, and use the training set to train the data center computing power load demand prediction model to obtain a trained data center computing power load demand prediction model. In the testing phase, the test set is input into the GA-PSO-BP network model; after reaching the maximum number of iterations, the predicted computing power load demand A is output; the computing power load demand A is then input into the GRU neural network to obtain a predicted computing power load demand data set B; after reaching the maximum number of iterations, the data set A is multiplied by the weight of the output result of the GA-PSO-BP network model to obtain a product A', and the data set B is multiplied by the weight of the output result of the GRU neural network to obtain a product B', and the sum of A' and B' is used as the predicted computing power load demand output, and the predicted computing power load demand output is evaluated to determine whether the trained data center computing power load demand prediction model meets the requirements;

[0226] The prediction module inputs the current workload data into the trained data center computing power load demand prediction model, and the data center computing power load demand prediction model outputs the computing power load demand of the data center in the future.

[0227] In another embodiment, the BP neural network optimization module further includes a parameter optimization module and a particle optimization module; the parameter optimization module is used to optimize the number of hidden layers of the BP network, the learning rate of the BP network, the node transfer function between the output layer and the hidden layer, the node transfer function of the output layer, and the training function; the particle optimization module is used to optimize the velocity v of particle i in the dth dimension (1≤d≤D) of the BP neural network. id and position X id Optimize, D is the total number of dimensions;

[0228] In another embodiment, the BP neural network optimization module further includes a particle crossover optimization module for optimizing the speed and position of the paired particles i and j in the crossover process of the BP neural network using the crossover operation in the GA algorithm.

[0229] In another embodiment, the BP neural network optimization module further includes a particle mutation optimization module for optimizing the speed and position of particles in the mutation process of the BP neural network by using the mutation operation in the GA algorithm.

[0230] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A combined prediction method for computing power load demand of a data center considering PUE, characterized in that: The steps include: S1. Establishing a data center computing load demand prediction model based on historical workload data generated by the GPU cluster within a preset time period. The data center computing load demand prediction model is a combined prediction model of a BP neural network and a GRU neural network optimized based on a GA-PSO optimization algorithm, and is used to predict computing load demand at future times based on current workload data. S2. Optimize the weights and thresholds of each dimension of the BP neural network using the GA-PSO optimization algorithm to obtain a GA-PSO-BP network model. The optimization objective is the inverse of the sum of squared sample errors, so as to minimize the fitting error of the data center computing power load demand prediction model. S3. Divide the historical data of the workload generated by the GPU cluster into a training set and a test set, use the training set to train the data center computing power load demand prediction model to obtain a trained data center computing power load demand prediction model, and in the testing phase, input the test set into the GA-PSO-BP network model; after reaching the maximum number of iterations, output the predicted computing power load demand A; then input the computing power load demand A into the GRU neural network to obtain a predicted computing power load demand data set B; after reaching the maximum number of iterations, multiply the data set A by the weight of the output result of the GA-PSO-BP network model to obtain a product A', multiply the data set B by the weight of the output result of the GRU neural network to obtain a product B', and use the sum of A' and B' as the predicted computing power load demand output, and evaluate the predicted computing power load demand output to determine whether the trained data center computing power load demand prediction model meets the requirements; S4. Input the current workload data into the data center computing power load demand prediction model trained in step S3. The data center computing power load demand prediction model outputs the computing power load demand of the data center at future times.

2. The data center computing power load demand combination prediction method considering PUE as claimed in claim 1, characterized in that: The method of optimizing the weights and thresholds of the BP neural network using the GA-PSO algorithm includes optimizing the following parameters: (1) Set the number of hidden layers of the BP network according to the following formula: Where: H is the number of hidden layer nodes; M is the number of nodes in the input layer; N is the number of nodes in the output layer; a is the set adjustment integer; (2) Set the learning rate η of the BP network according to the following formula: Where: T now is the current iteration number; T max is the maximum number of iterations; (3) Set the tansig function as the node transfer function between the output layer and the hidden layer; the purelin linear function as the node transfer function of the output layer; and the trainlm function as the training function.

3. The data center computing power load demand combination prediction method considering PUE as claimed in claim 1, characterized in that: The method of using GA-PSO algorithm to optimize the weights and thresholds of BP neural network also includes the velocity v of particle i in the dth dimension in BP neural network. id and position X id Optimize The velocity v of particle i in the dth dimension id and position X id The update formula with individual extreme value and group extreme value is as follows: X id =X id +v id Where: V id : represents the velocity of the i-th particle in the d-th dimension; X id : represents the position of the i-th particle in the d-th dimension; t is the current iteration number; d is the spatial dimension of the solution, 1≤d≤D, D is the total dimension; ω is the inertia weight; C1 and C2 are learning factors; Represents a random number in the interval [0,1]; Pbest id is the individual extreme value of the i-th particle in the d-th dimension; Gbest id is the group extreme value of the i-th particle in the d-th dimension.

4. The method for predicting data center computing power load demand combination considering PUE according to claim 1, characterized in that: The method of using GA-PSO algorithm to optimize the weights and thresholds of BP neural network also includes using the crossover operation in GA algorithm to optimize the speed and position of paired particles i and j in the crossover process of BP neural network. The formula is as follows: Speed ​​Cross: Position Crossover: Where: α1, α2 are random numbers within the set interval; V id (t), V jd (t): represents the velocity of the i-th particle and the j-th particle at time t in the d-th dimension respectively; X id (t) and X jd (t): represents the position of the i-th particle and the j-th particle at time t in the d-th dimension; d is the spatial dimension of the solution, 1≤d≤D, and D is the total dimension.

5. The method for predicting data center computing power load demand combination considering PUE according to claim 1, characterized in that: The method of using GA-PSO algorithm to optimize the weights and thresholds of BP neural network also includes using the mutation operation in GA algorithm to optimize the speed and position of particle i in the mutation process of BP neural network. The formula is as follows: Speed ​​variation: Position variation: Where: V id (t): represents the velocity of the i-th particle at time t in the d-th dimension; X id (t): represents the position of the i-th particle at time t in the d-th dimension; V min , V max : are the minimum and maximum velocities of particle i respectively; X min , X max : are the lower and upper bounds of the position of particle i respectively; d is the spatial dimension of the solution, 1≤d≤D, D is the total dimension; is a random number in the interval [0,1].

6. The method for predicting data center computing power load demand combination considering PUE according to claim 1, characterized in that: The method of using the GA-PSO algorithm to optimize the weights and thresholds of the BP neural network also includes optimizing the fitness function in the BP neural network, and its calculation formula is: Where: fitness now : represents the updated fitness value; ω now : represents the updated inertia weight; fitness past : Indicates the fitness value before updating; ω past and ω now : represent the inertia weights before and after iteration respectively, and ω past and ω now The sum of is the set value; A and B are both set constants; p i is the network prediction value of the i-th training sample; r i is the actual output value of the i-th training sample; N is the total number of training samples.

7. A data center computing power load demand combination prediction system considering PUE according to the method of claim 1, characterized in that: It includes a data center computing power load demand prediction model building module, a BP neural network optimization module, a model training module, and a prediction module; The data center computing power load demand prediction model construction module is used to establish a data center computing power load demand prediction model based on the historical data of the workload generated by the GPU cluster within a preset time period. The data center computing power load demand prediction model is a combined prediction model of the BP neural network and the GRU neural network optimized based on the GA-PSO optimization algorithm, and is used to predict the computing power load demand at future moments based on the workload data at the current moment; The BP neural network optimization module is used to optimize the weights and thresholds of each dimension of the BP neural network using the GA-PSO optimization algorithm to obtain a GA-PSO-BP network model. Its optimization objective is the inverse of the sum of squared sample errors, so as to minimize the fitting error of the data center computing power load demand prediction model. The model training module is used to divide the historical data of the workload generated by the GPU cluster into a training set and a test set, and use the training set to train the data center computing power load demand prediction model to obtain a trained data center computing power load demand prediction model. In the testing phase, the test set is input into the GA-PSO-BP network model; after reaching the maximum number of iterations, the predicted computing power load demand A is output; the computing power load demand A is then input into the GRU neural network to obtain a predicted computing power load demand data set B; after reaching the maximum number of iterations, the data set A is multiplied by the weight of the output result of the GA-PSO-BP network model to obtain a product A', and the data set B is multiplied by the weight of the output result of the GRU neural network to obtain a product B', and the sum of A' and B' is used as the predicted computing power load demand output, and the predicted computing power load demand output is evaluated to determine whether the trained data center computing power load demand prediction model meets the requirements; The prediction module inputs the current workload data into the trained data center computing power load demand prediction model, and the data center computing power load demand prediction model outputs the computing power load demand of the data center at future times.

8. The data center computing power load demand combination prediction system considering PUE as claimed in claim 7, characterized in that: The BP neural network optimization module also includes a parameter optimization module and a particle optimization module; The parameter optimization module is used to optimize the number of hidden layers of the BP network, the learning rate of the BP network, the node transfer function between the output layer and the hidden layer, the node transfer function of the output layer, and the training function; The particle optimization module is used to calculate the velocity v of particle i in the dth dimension in the BP neural network. id and position X id Optimize, where 1≤d≤D, D is the total number of dimensions.

9. The data center computing power load demand combination prediction system considering PUE as claimed in claim 7, characterized in that: The BP neural network optimization module also includes a particle crossover optimization module for optimizing the speed and position of the paired particles i and j in the crossover process of the BP neural network by using the crossover operation in the GA algorithm.

10. The data center computing power load demand combination prediction system considering PUE according to claim 7, characterized in that: The BP neural network optimization module also includes a particle mutation optimization module for optimizing the speed and position of particles in the mutation process of the BP neural network by using the mutation operation in the GA algorithm.

Images