Energy cost prediction method for computing infrastructure based on deep learning

Abstract

The present invention discloses a method for predicting the energy cost of computing power infrastructure based on deep learning, and the method includes data collection, sub-model training, establishing a computing power infrastructure energy cost prediction model and computing power infrastructure energy cost prediction. The present invention belongs to the field of cost prediction technology, and specifically refers to a method for predicting the energy cost of computing power infrastructure based on deep learning. This scheme adopts dynamic feature selection, LSTM-based sub-model training, data set distance design and undersampling strategy, which can better adapt to the ever-changing energy mode, improve the ability to capture the characteristics of complex facility energy data, and thus ensure the accuracy of the final energy cost prediction; by evaluating the number of incorrect predictions of each base classifier, dynamically adjusting the weight of the base classifier, and finally integrating the prediction results, the accuracy of the prediction results is improved; the model is quickly converged through a unique initial parameter adjustment strategy, thereby improving the efficiency of cost prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of cost prediction, and specifically to a method for predicting energy cost of computing power infrastructure based on deep learning. Background Art

[0002] The energy cost prediction method for computing power infrastructure is a method used to evaluate and predict the energy cost of computing power infrastructure such as data centers or cloud computing platforms, aiming to provide enterprises with accurate energy cost prediction and management solutions. However, the general energy cost prediction method for computing power infrastructure has the problems of insufficient dynamic adaptability of the original facility energy data, improper feature selection, and low attention to important samples, which leads to low cost prediction accuracy; the general energy cost prediction method for computing power infrastructure has the problems of inability to respond to input features in a timely manner, weak model learning ability, and slow convergence of model parameters, which leads to insufficient convergence of cost prediction and low efficiency. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a computing power infrastructure energy cost prediction method based on deep learning. The general computing power infrastructure energy cost prediction method has the problems of insufficient dynamic adaptability of the original facility energy data, improper feature selection, and low attention to important samples, which leads to low cost prediction accuracy. This solution adopts dynamic feature selection, LSTM-based sub-model training, data set distance design and undersampling strategy, which can better adapt to the changing energy mode, improve the ability to capture complex facility energy data characteristics, and provide more reliable and accurate support for energy cost prediction, thereby ensuring the accuracy of the final energy cost prediction; in view of the general computing power infrastructure energy cost prediction method, the inability to respond to input features in a timely manner, weak model learning ability, and slow model parameter adjustment convergence speed, which leads to insufficient convergence of cost prediction and low efficiency, this solution evaluates the number of incorrect predictions of each base classifier, dynamically adjusts the weight of the base classifier, and finally integrates the prediction results to achieve rapid learning of the model, thereby improving the accuracy of the prediction results; a unique initial parameter adjustment strategy is used to achieve rapid convergence of the model, thereby improving the efficiency of cost prediction.

[0004] The technical solution adopted by the present invention is as follows: The method for predicting energy cost of computing power infrastructure based on deep learning provided by the present invention comprises the following steps:

[0005] Step S1: data collection;

[0006] Step S2: sub-model training;

[0007] Step S3: Establishing a computing power infrastructure energy cost prediction model;

[0008] Step S4: Prediction of computing infrastructure energy costs.

[0009] Furthermore, in step S1, the data collection is to collect historical data of computing power infrastructure; the historical data of computing power infrastructure includes equipment power consumption, equipment operating time, load, temperature, humidity, external temperature, equipment status monitoring data, cooling system performance and historical energy cost; the historical energy cost is used as a data label; the data set is divided after data cleaning and data conversion to obtain a training set and a test set.

[0010] Furthermore, in step S2, the sub-model training specifically includes the following steps:

[0011] Step S21: Generate a sub-dataset; introduce dynamic feature selection, dynamically adjust feature weights according to energy consumption in different time periods, and use sliding window technology to adjust features in real time. The dynamic feature weight is expressed as: ; Generate L sub-datasets, each of which uses a different feature combination and time window; use each sub-dataset to train a base classifier; use LSTM as the base classifier, and use the cross entropy loss function as the loss function; for each base classifier, dynamically update the parameters, expressed as: ;in, is the dynamic feature weight; n is the total number of samples; i is the sample index; y i is the actual energy cost; is to predict energy costs; is the characteristic value of the sample; are the updated base classifier parameters; is the base classifier parameter before updating; is the learning rate; is the loss function of the base classifier hl relative to the parameter The gradient of is the regulating factor; is the loss function of the integrated model relative to the parameter The gradient of is the regularization parameter;

[0012] Step S22: Design the data set distance; expressed as follows:

[0013] ;

[0014] In the formula, is the dataset distance between sub-datasets p1 and q; m is the total number of features, k is the feature index; and are the k-th eigenvalues ​​of the i1-th sample and the i2-th sample in the sub-dataset respectively;

[0015] Step S23: Overall dataset distance between sub-datasets; the overall SJJ is calculated by averaging the pairwise diversity in all L sub-datasets, using the following formula:

[0016] ;

[0017] Where SJJ is the overall dataset distance between sub-datasets;

[0018] Step S24: undersampling; train a base classifier for each generated sub-data set, and generate multiple models through the bagging strategy. Bagging is performed by resampling the input data, and in each round of training, samples with large prediction residuals are preferentially sampled. The prediction residual is expressed as: ; Then the output results of all base classifiers are averaged or voted to generate the final result; where, is the sampling probability of the i-th sample; are the prediction residuals.

[0019] Furthermore, in step S3, establishing a computing power infrastructure energy cost prediction model specifically includes the following steps:

[0020] Step S31: Calculate the prediction score; for the lth classifier of the i-th instance, the prediction score is expressed as: ; and provide prediction uncertainty for the base classifier based on the Bayesian neural network, expressed as: ; Where zli is the predicted score of the sample; is the prediction score of the lth base classifier for the i-th sample; is the prediction uncertainty of the l-th base classifier for the i-th sample; is the input feature of the i-th sample; is the parameter of the lth base classifier;

[0021] Step S32: Calculate the number of incorrect predictions of the base classifier; for each sample, the number of incorrect predictions is expressed as: ; When zli=1, =0, when zli=-1, =1; where is the number of incorrect predictions of samples on all base classifiers;

[0022] Step S33: Calculate the base classifier score; the base classifier score is adaptively calculated based on its prediction result for each instance, expressed as: ; After normalization, the final base classifier score SL is obtained; where sli is the score of the l-th classifier on the i-th sample;

[0023] Step S34: Calculate voting weights; assign initial weights wl to each base classifier, expressed as:

[0024] ;in, ; is the initial base classifier weight; the weight is dynamically adjusted at each iteration, expressed as: ; is the updated base classifier weight; α1 is the weight adjustment step; The base classifier hl is based on the new sample (x new ,y new )’s prediction residuals;

[0025] Step S35: Calculate the integrated prediction result; predict the energy cost based on the weighted voting of each model, expressed as: ; Where y(x) is the prediction result of establishing the computing power infrastructure energy cost prediction model;

[0026] Step S36: training judgment; when the computing power infrastructure energy cost prediction model converges with the training set loss, the computing power infrastructure energy cost prediction model training is completed; a prediction threshold is set in advance, and when the prediction accuracy of the trained computing power infrastructure energy cost prediction model for the test set is higher than the prediction threshold, the computing power infrastructure energy cost prediction model is established; otherwise, go to step S37 to adjust the initial parameters and re-divide the data set training model;

[0027] Step S37: initial parameter adjustment; specifically including:

[0028] Step S371: Initialization; construct parameter adjustment space based on learning rate, initial base classifier weight, adjustment factor, regularization parameter and initial parameters of base classifier; initialize and optimize individual positions of population; use the prediction accuracy of the computing power infrastructure energy cost prediction model trained based on individual positions for the test set as the individual fitness value; the individual with the highest individual fitness value is regarded as the optimal individual; the formula used to initialize and optimize individual positions of population is as follows:

[0029] ;

[0030] In the formula, x n3+1,j3 and x n3,j3 are the initialization positions of the j3th dimension of the n3+1th individual and the n3th individual respectively; p is the adjustment parameter;

[0031] Step S372: Optimize the position in the exploration phase; the formula used is as follows:

[0032] ;

[0033] ;

[0034] In the formula, x i3,j3 (·) is the position of the j3th dimension of the i3th individual, t is the current search number; B j3 (·) is the optimal individual position; β1 is the adjustment factor; R j3 (·) is the worst position; rand is a random number between 0 and 1; is the movement weight factor, is the initial moving weight factor; T is the maximum number of searches; x r1,j3 (·) is the random individual position;

[0035] Step S373: Position optimization in the development phase; the formula used is as follows:

[0036] ;

[0037] Where P i3,j3 (·) is the optimal position of individual search experience; is the reverse factor;

[0038] Step S374: random disturbance; a disturbance threshold is set in advance, and random disturbance is applied to individuals with low fitness values ​​whose fitness values ​​differ from each other by less than the disturbance threshold. The formula used is as follows:

[0039] ;

[0040] Where n3ewx i3,j3 (·) is the position after disturbance; is the disturbance factor;

[0041] Step S375: Jump strategy; a jump threshold is preset. When the increase in the position fitness value after the individual update is higher than the jump threshold, the jump strategy is used for the individual. The formula used is as follows:

[0042] ;

[0043] Where n3ew1x i3,j3 (·) is the position after the jump; J is the jump amplitude;

[0044] Step S376: search judgment; when there is an individual fitness value higher than the prediction threshold, the computing power infrastructure energy cost prediction model obtained based on the individual position is established; if the maximum number of searches is reached, go to step S371; otherwise, go to step S372.

[0045] Furthermore, in step S4, the computing power infrastructure energy cost prediction is based on the established computing power infrastructure energy cost prediction model, and real-time collection of equipment power consumption, equipment operating time, load, temperature, humidity, external temperature, equipment status monitoring data and cooling system performance are performed, and input into the computing power infrastructure energy cost prediction model after preprocessing, and the historical energy cost predicted by the model is used as the energy cost prediction result.

[0046] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0047] (1) In view of the problems of insufficient dynamic adaptability of the original facility energy data, improper feature selection, and low attention to important samples in general computing infrastructure energy cost prediction methods, which lead to low cost prediction accuracy, this scheme adopts dynamic feature selection, LSTM-based sub-model training, data set distance design and undersampling strategy, which can better adapt to the changing energy pattern, improve the ability to capture the characteristics of complex facility energy data, provide more reliable and accurate support for energy cost prediction, and thus ensure the accuracy of the final energy cost prediction.

[0048] (2) To address the problems of general computing infrastructure energy cost prediction methods, such as the inability to respond to input features in a timely manner, weak model learning ability, and slow model parameter convergence, which in turn lead to insufficient convergence of cost prediction and low efficiency, this solution evaluates the number of incorrect predictions of each base classifier, dynamically adjusts the weights of the base classifiers, and finally integrates the prediction results to achieve rapid model learning, thereby improving the accuracy of the prediction results; and achieves rapid model convergence through a unique initial parameter adjustment strategy, thereby improving the efficiency of cost prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A schematic diagram of a process flow of a computing infrastructure energy cost prediction method based on deep learning provided by the present invention;

[0050] Figure 2 is a schematic flow chart of step S2;

[0051] Figure 3 It is a schematic diagram of the process of step S3.

[0052] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0055] Example 1, see Figure 1 The present invention provides a method for predicting energy cost of computing power infrastructure based on deep learning, and the method comprises the following steps:

[0056] Step S1: Data collection, collecting historical data of computing power infrastructure;

[0057] Step S2: sub-model training, based on dynamic feature selection and data set distance design, to achieve sub-model training;

[0058] Step S3: Establish a computing power infrastructure energy cost prediction model, evaluate the number of incorrect predictions of each base classifier, dynamically adjust the weights of the base classifiers, and finally integrate the prediction results. Based on the initial parameter adjustment strategy, the computing power infrastructure energy cost prediction model is established;

[0059] Step S4: Computing power infrastructure energy cost prediction, realizing cost prediction based on the established computing power infrastructure energy cost prediction model.

[0060] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the historical data of computing infrastructure includes device power consumption, device operating time, load, temperature, humidity, external temperature, device status monitoring data, cooling system performance and historical energy cost; the historical energy cost is used as a data label; the data set is divided after data cleaning and data conversion to obtain a training set and a test set.

[0061] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, the sub-model training specifically includes the following steps:

[0062] Step S21: Generate a sub-dataset; introduce dynamic feature selection, dynamically adjust feature weights according to energy consumption in different time periods, and use sliding window technology to adjust features in real time. The dynamic feature weight is expressed as: ; Generate L sub-datasets, each of which uses a different feature combination and time window; use each sub-dataset to train a base classifier; use LSTM as the base classifier, and use the cross entropy loss function as the loss function; for each base classifier, dynamically update the parameters, expressed as: ;in, is the dynamic feature weight; n is the total number of samples; i is the sample index; y i is the actual energy cost; is to predict energy costs; is the characteristic value of the sample; are the updated base classifier parameters; is the base classifier parameter before updating; is the learning rate; is the loss function of the base classifier hl relative to the parameter The gradient of is the regulating factor; is the loss function of the integrated model relative to the parameter The gradient of is the regularization parameter;

[0063] Step S22: Design the data set distance; expressed as follows:

[0064] ;

[0065] In the formula, is the dataset distance between sub-datasets p1 and q; m is the total number of features, k is the feature index; and are the k-th eigenvalues ​​of the i1-th sample and the i2-th sample in the sub-dataset respectively;

[0066] Step S23: Overall dataset distance between sub-datasets; the overall SJJ is calculated by averaging the pairwise diversity in all L sub-datasets, using the following formula:

[0067] ;

[0068] Where SJJ is the overall dataset distance between sub-datasets;

[0069] Step S24: undersampling; train a base classifier for each generated sub-data set, and generate multiple models through the bagging strategy. Bagging is performed by resampling the input data, and in each round of training, samples with large prediction residuals are preferentially sampled. The prediction residual is expressed as: ; Then the output results of all base classifiers are averaged or voted to generate the final result; where, is the sampling probability of the i-th sample; are the prediction residuals.

[0070] By performing the above operations, the general computing infrastructure energy cost prediction method has the problems of insufficient dynamic adaptability of original facility energy data, improper feature selection, and low attention to important samples, which leads to low cost prediction accuracy. This solution adopts dynamic feature selection, LSTM-based sub-model training, data set distance design and undersampling strategy, which can better adapt to the changing energy patterns, improve the ability to capture the characteristics of complex facility energy data, provide more reliable and accurate support for energy cost prediction, and thus ensure the accuracy of the final energy cost prediction.

[0071] Example 4, see Figure 1 and Figure 3 Based on the above embodiment, in step S3, establishing a computing power infrastructure energy cost prediction model specifically includes the following steps:

[0072] Step S31: Calculate the prediction score; for the lth classifier of the i-th instance, the prediction score is expressed as: ; and provide prediction uncertainty for the base classifier based on the Bayesian neural network, expressed as: ; Where zli is the predicted score of the sample; is the prediction score of the lth base classifier for the i-th sample; is the prediction uncertainty of the l-th base classifier for the i-th sample; is the input feature of the i-th sample; is the parameter of the lth base classifier;

[0073] Step S32: Calculate the number of incorrect predictions of the base classifier; for each sample, the number of incorrect predictions is expressed as: ; When zli=1, =0, when zli=-1, =1; where is the number of incorrect predictions of samples on all base classifiers;

[0074] Step S33: Calculate the base classifier score; the base classifier score is adaptively calculated based on its prediction result for each instance, expressed as: ; After normalization, the final base classifier score SL is obtained; where sli is the score of the l-th classifier on the i-th sample;

[0075] Step S34: Calculate the voting weight; assign an initial weight wl to each base classifier, expressed as: ;in, ; is the initial base classifier weight; the weight is dynamically adjusted at each iteration, expressed as: ; is the updated base classifier weight; α1 is the weight adjustment step; The base classifier hl is based on the new sample (x new ,y new )’s prediction residuals;

[0076] Step S35: Calculate the integrated prediction result; predict the energy cost based on the weighted voting of each model, expressed as: ; Where y(x) is the prediction result of establishing the computing power infrastructure energy cost prediction model;

[0077] Step S36: training judgment; when the computing power infrastructure energy cost prediction model converges with the training set loss, the computing power infrastructure energy cost prediction model training is completed; a prediction threshold is set in advance, and when the prediction accuracy of the trained computing power infrastructure energy cost prediction model for the test set is higher than the prediction threshold, the computing power infrastructure energy cost prediction model is established; otherwise, go to step S37 to adjust the initial parameters and re-divide the data set training model;

[0078] Step S37: initial parameter adjustment; specifically including:

[0079] Step S371: Initialization; construct parameter adjustment space based on learning rate, initial base classifier weight, adjustment factor, regularization parameter and initial parameters of base classifier; initialize and optimize individual positions of population; use the prediction accuracy of the computing power infrastructure energy cost prediction model trained based on individual positions for the test set as the individual fitness value; the individual with the highest individual fitness value is regarded as the optimal individual; the formula used to initialize and optimize individual positions of population is as follows:

[0080] ;

[0081] In the formula, x n3+1,j3 and x n3,j3 are the initialization positions of the j3th dimension of the n3+1th individual and the n3th individual respectively; p is the adjustment parameter;

[0082] Step S372: Exploration phase position optimization; the formula used is as follows:

[0083] ;

[0084] ;

[0085] In the formula, x i3,j3 (·) is the position of the j3th dimension of the i3th individual, t is the current search number; B j3 (·) is the optimal individual position; β1 is the adjustment factor; R j3 (·) is the worst position; rand is a random number between 0 and 1; is the movement weight factor, is the initial moving weight factor; T is the maximum number of searches; x r1,j3 (·) is the random individual position;

[0086] Step S373: Position optimization in the development phase; the formula used is as follows:

[0087] ;

[0088] Where P i3,j3 (·) is the optimal position of individual search experience; is the reverse factor;

[0089] Step S374: random disturbance; a disturbance threshold is set in advance, and random disturbance is applied to individuals with low fitness values ​​whose fitness values ​​differ from each other by less than the disturbance threshold. The formula used is as follows:

[0090] ;

[0091] Where n3ewx i3,j3 (·) is the position after disturbance; is the disturbance factor;

[0092] Step S375: Jump strategy; a jump threshold is preset. When the increase in the position fitness value after the individual update is higher than the jump threshold, the jump strategy is used for the individual. The formula used is as follows:

[0093] ;

[0094] Where n3ew1x i3,j3 (·) is the position after the jump; J is the jump amplitude;

[0095] Step S376: search judgment; when there is an individual fitness value higher than the prediction threshold, the computing power infrastructure energy cost prediction model obtained based on the individual position is established; if the maximum number of searches is reached, go to step S371; otherwise, go to step S372.

[0096] By performing the above operations, this solution can address the problems of general computing infrastructure energy cost prediction methods, such as the inability to respond to input features in a timely manner, weak model learning ability, and slow model parameter convergence, which in turn leads to insufficient convergence of cost prediction and low efficiency. By evaluating the number of incorrect predictions of each base classifier, dynamically adjusting the weights of the base classifiers, and finally integrating the prediction results, this solution can achieve rapid model learning and thus improve the accuracy of the prediction results. A unique initial parameter adjustment strategy can be used to achieve rapid model convergence and thus improve the efficiency of cost prediction.

[0097] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the computing power infrastructure energy cost prediction is based on the established computing power infrastructure energy cost prediction model. The equipment power consumption, equipment operating time, load, temperature, humidity, external temperature, equipment status monitoring data and cooling system performance are collected in real time, and input into the computing power infrastructure energy cost prediction model after preprocessing, and the historical energy cost predicted by the model is used as the energy cost prediction result.

[0098] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0099] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0100] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A computing infrastructure energy cost prediction method based on deep learning, characterized by: The method comprises the following steps: Step S1: Data collection, collecting historical data of computing power infrastructure; Step S2: sub-model training, based on dynamic feature selection and sample distance design, to achieve sub-model training; Step S3: Establish a computing power infrastructure energy cost prediction model, evaluate the number of incorrect predictions of each base classifier, dynamically adjust the weights of the base classifiers, and finally integrate the prediction results. Based on the initial parameter adjustment strategy, the computing power infrastructure energy cost prediction model is established; Step S4: computing power infrastructure energy cost prediction, realizing cost prediction based on the established computing power infrastructure energy cost prediction model; Step S2 includes step S22: designing the data set distance; which is expressed as follows: ； In the formula, is the dataset distance between sub-datasets p1 and q; m is the total number of features, k is the feature index; and are the k-th eigenvalues ​​of the i1-th sample and the i2-th sample in the sub-dataset respectively; In step S2, the sub-model training further includes the following steps: Step S21: Generate a sub-dataset; introduce dynamic feature selection, dynamically adjust feature weights according to energy consumption in different time periods, and use sliding window technology to adjust features in real time. The dynamic feature weight is expressed as: ; Generate L sub-datasets, each of which uses a different feature combination and time window; use each sub-dataset to train a base classifier; use LSTM as the base classifier, and use the cross entropy loss function as the loss function; for each base classifier, dynamically update the parameters, expressed as: ;in, is the dynamic feature weight; n is the total number of samples; i is the sample index; y i is the actual energy cost; is to predict energy costs; is the characteristic value of the sample; are the updated base classifier parameters; is the base classifier parameter before updating; is the learning rate; is the loss function of the base classifier hl relative to the parameter The gradient of is the regulating factor; is the loss function of the integrated model relative to the parameter The gradient of is the regularization parameter; Step S23: Overall dataset distance between sub-datasets; the overall SJJ is calculated by averaging the pairwise diversity in all L sub-datasets, using the following formula: ； Where SJJ is the overall dataset distance between sub-datasets; Step S24: undersampling; train a base classifier for each generated sub-data set, and generate multiple models through the bagging strategy. Bagging is performed by resampling the input data, and in each round of training, samples with large prediction residuals are preferentially sampled. The prediction residual is expressed as: ; Then the output results of all base classifiers are averaged or voted to generate the final result; where, is the sampling probability of the i-th sample; are the prediction residuals.

2. The method for predicting computing infrastructure energy costs based on deep learning according to claim 1, characterized in that: In step S3, the establishment of a computing power infrastructure energy cost prediction model specifically includes the following steps: Step S31: Calculate the prediction score; for the lth classifier of the i-th instance, the prediction score is expressed as: ; and provide prediction uncertainty for the base classifier based on the Bayesian neural network, expressed as: ; Where zli is the predicted score of the sample; is the prediction score of the lth base classifier for the i-th sample; is the prediction uncertainty of the l-th base classifier for the i-th sample; is the input feature of the i-th sample; is the parameter of the lth base classifier; Step S32: Calculate the number of incorrect predictions of the base classifier; for each sample, the number of incorrect predictions is expressed as: ; When zli=1, =0, when zli=-1, =1; where is the number of incorrect predictions of samples on all base classifiers; Step S33: Calculate the base classifier score; the base classifier score is adaptively calculated based on its prediction result for each instance, expressed as: ; After normalization, the final base classifier score SL is obtained; where sli is the score of the l-th classifier on the i-th sample; Step S34: Calculate voting weights; assign initial weights wl to each base classifier, expressed as: ;in, ; is the initial base classifier weight; the weight is dynamically adjusted at each iteration, expressed as: ; is the updated base classifier weight; α1 is the weight adjustment step; The base classifier hl is based on the new sample (x new ,y new )’s prediction residuals; Step S35: Calculate the integrated prediction result; predict the energy cost based on the weighted voting of each model, expressed as: ; Where y(x) is the prediction result of establishing the computing power infrastructure energy cost prediction model; Step S36: training judgment; when the computing power infrastructure energy cost prediction model converges with the training set loss, the computing power infrastructure energy cost prediction model training is completed; a prediction threshold is set in advance, and when the prediction accuracy of the trained computing power infrastructure energy cost prediction model for the test set is higher than the prediction threshold, the computing power infrastructure energy cost prediction model is established; otherwise, go to step S37 to adjust the initial parameters and re-divide the data set training model; Step S37: initial parameter adjustment.

3. The method for predicting computing power infrastructure energy cost based on deep learning according to claim 2, characterized in that: In step S37, the initial parameter adjustment specifically includes: Step S371: Initialization; construct parameter adjustment space based on learning rate, initial base classifier weight, adjustment factor, regularization parameter and initial parameters of base classifier; initialize and optimize individual positions of population; use the prediction accuracy of the computing power infrastructure energy cost prediction model trained based on individual positions for the test set as the individual fitness value; the individual with the highest individual fitness value is regarded as the optimal individual; the formula used to initialize and optimize individual positions of population is as follows: ； In the formula, x n3+1,j3 and x n3,j3 are the initialization positions of the j3th dimension of the n3+1th individual and the n3th individual respectively; p is the adjustment parameter; Step S372: Optimize the position in the exploration phase; the formula used is as follows: ； ； In the formula, x i3,j3 (·) is the position of the j3th dimension of the i3th individual, t is the current search number; B j3 (·) is the optimal individual position; β1 is the adjustment factor; R j3 (·) is the worst position; rand is a random number between 0 and 1; is the movement weight factor, is the initial moving weight factor; T is the maximum number of searches; x r1,j3 (·) is the random individual position; Step S373: Position optimization in the development phase; the formula used is as follows: ； Where P i3,j3 (·) is the optimal position of individual search experience; is the reverse factor; Step S374: random disturbance; a disturbance threshold is set in advance, and random disturbance is applied to individuals with low fitness values ​​whose fitness values ​​differ from each other by less than the disturbance threshold. The formula used is as follows: ； Where n3ewx i3,j3 (·) is the position after disturbance; is the disturbance factor; Step S375: Jump strategy; a jump threshold is preset. When the increase in the position fitness value after the individual update is higher than the jump threshold, the jump strategy is used for the individual. The formula used is as follows: ； Where n3ew1x i3,j3 (·) is the position after the jump; J is the jump amplitude; Step S376: search judgment; when there is an individual fitness value higher than the prediction threshold, the computing power infrastructure energy cost prediction model obtained based on the individual position is established; if the maximum number of searches is reached, go to step S371; otherwise, go to step S372.

4. The method for predicting computing power infrastructure energy cost based on deep learning according to claim 1, characterized in that: In step S1, the historical data of the computing infrastructure includes device power consumption, device operating time, load, temperature, humidity, external temperature, device status monitoring data, cooling system performance and historical energy cost; the historical energy cost is used as a data label; the data set is cleaned and converted, and then the data set is divided to obtain a training set and a test set.

5. The method for predicting computing infrastructure energy costs based on deep learning according to claim 1, characterized in that: In step S4, the computing power infrastructure energy cost prediction is based on the established computing power infrastructure energy cost prediction model, and real-time data such as equipment power consumption, equipment operating time, load, temperature, humidity, external temperature, equipment status monitoring data and cooling system performance are collected. After preprocessing, the data is input into the computing power infrastructure energy cost prediction model, and the historical energy cost predicted by the model is used as the energy cost prediction result.

Images