CN110852498A - A method for predicting data center energy efficiency value PUE based on GRU neural network - Google Patents

Abstract

The invention discloses a method for predicting the energy consumption efficiency value PUE of a data center based on a GRU neural network, comprising the following steps: 1) collecting attribute data related to the energy consumption of the data center; 2) normalizing the attribute data collected in step 1). Normalization and feature selection; 3) Select the GRU prediction model as the prediction model, set the loss function J(w) and the optimization method optimizer, and then use part of the data obtained in step 2) feature selection to train the GRU prediction model; 4) Utilize the steps 2) The remaining data obtained by feature selection is used to evaluate the trained GRU prediction model, and then use the evaluated GRU prediction model to predict the data center energy consumption efficiency value PUE. This method can more accurately predict the data center energy consumption efficiency value PUE.

Description

A method for predicting the energy efficiency value PUE of data center based on GRU neural network

technical field

The invention belongs to the technical field of data center energy saving, and relates to a method for predicting a data center energy consumption efficiency value PUE based on a GRU neural network.

Background technique

With the rapid development of technologies such as cloud computing, the Internet of Things, and artificial intelligence, the scale and number of data centers as infrastructure have grown rapidly. A data center is a collection of large-scale electrical equipment, including IT equipment for processing, storing, and forwarding data, cooling control systems and power supply systems that maintain the environment at a suitable temperature and humidity. In order to ensure the normal operation of these equipment. Operation, its power consumption is very huge. In 2014, U.S. data centers used approximately 70 billion kWh of electricity, or approximately 1.8% of total U.S. electricity consumption, and based on current trends, data centers in the U.S. are expected to consume approximately 73 billion kWh by 2020 . Due to the consideration of operating cost, energy, environment, etc., reducing the power consumption of the data center and improving the energy consumption efficiency of the data center are urgent problems to be solved at present.

The energy efficiency of data centers usually uses PUE (Power Usage Effectiveness) as an evaluation criterion. PUE represents the ratio of the total power supplied to the data center to the power consumption only used to supply IT equipment. In theory, the closer the PUE is to 1, the higher the energy consumption efficiency. In the energy management process of the data center, PUE can not only be used to evaluate the energy consumption efficiency of the data center, but also provide relevant information such as power demand for the energy management of the data center. Therefore, if the PUE of the data center can be accurately predicted, it will provide effective suggestions for the energy management of the data center. However, the energy consumption composition of a data center is very complex. Servers, cooling systems, power supply systems, weather and environment all affect energy consumption, so it is very challenging to accurately predict PUE.

Among the existing methods for predicting PUE and energy consumption, Google proposed the use of ordinary ANN (Artificial Neural Network) to predict the PUE of the data center, and some other studies proposed the use of an expert system with confidence (ABelief Rule Based Expert System) , polynomial linear regression model and other methods to predict the PUE of the data center. The above works have opened up new ideas for data center energy consumption prediction, but there are also some shortcomings. On the one hand, the attributes considered for predicting PUE are not strongly or comprehensively related to PUE; on the other hand, the model used does not have the function of considering the time series of energy consumption attributes of the data center, and uses non-time series For related machine learning algorithms, this ignores the characteristics of continuous changes in some time series variables such as temperature and humidity.

Therefore, the premise of accurately predicting data center PUE is to fully consider as many features as possible related to energy consumption and the temporal characteristics of these features.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art, and provide a method for predicting the energy consumption efficiency value PUE of a data center based on a GRU neural network, which can more accurately predict the energy consumption efficiency value PUE of the data center.

In order to achieve the above object, the method for predicting the energy consumption efficiency value PUE of a data center based on a GRU neural network according to the present invention comprises the following steps:

1) Collect attribute data related to data center energy consumption;

2) Normalization and feature selection are performed on the attribute data collected in step 1);

3) Select the GRU prediction model as the prediction model, set the loss function J(w) and the optimization method optimizer, and then use the part of the data obtained by the feature selection in step 2) to train the GRU prediction model;

4) Evaluate the trained GRU prediction model using the remaining data obtained from the feature selection in step 2), and then use the evaluated GRU prediction model to predict the energy consumption efficiency value PUE of the data center.

The specific operation of step 1) is: collecting the attribute data related to the energy consumption of the data center at a set time interval, wherein the collected attribute data related to the energy consumption of the data center includes attribute data related to IT equipment, attribute data related to the environment, Attribute data related to cooling system and attribute data related to infrastructure, among which, attribute data related to IT equipment includes server load, UPS power load and heat increase rate; environment-related attribute data includes temperature, humidity and dew point temperature; cooling system Relevant attribute data includes cooling rate and CRAC power; infrastructure-related attribute data includes lighting and HVAC power. Set the attribute data related to data center energy consumption collected at time t X _t = (x ¹ , x ² , .. ., ^xn ).

The specific operations of step 2) are:

21) Normalize the collected attribute data X _t =(x ¹ ,x ² ,...,x ⁿ ) related to energy consumption of the data center to [0,1], and normalize the normalized data Denoted as X _i =(x ¹ ,x ² ,...,x ⁿ );

22) Use the feature selection algorithm to select m attribute data with the greatest correlation with the predicted PUE.

The specific operation of step 22) is:

Feature selection is performed using the RFECV method, and m attribute data X=(x ¹ , x ² , . . . , x ^m ) that are most relevant to the predicted PUE are selected.

In step 3), the GRU recurrent neural network is used as the PUE prediction model, wherein, in the GRU neuron structure in the GRU recurrent neural network, there are two gate control units, a reset gate r _t and an update gate z _t , which are used to realize memory or Forget the training information generated by the previous training step, the specific calculation process of the GRU unit is:

z _t =σ(W _z ·[h _t-1 , x _t ]+b _z )

r _t =σ(W _r ·[h _t-1 , x _t ]+ _br )

Among them, z _t is used to control how much information h _t-1 of the previous state is written to the current state h _t , σ( ) is the sigmoid activation function that scales the value between (0, 1), namely:

t is the incoming parameter to be scaled, where the closer the value of σ( ) is to 0, the more writing, the closer the value of σ( ) is to 1, the less the writing is, where h _{t- 1} represents the state information output at time t-1, x _t represents the input at the current time, W _z , W _r and W _n are weight parameters, and b _z , _br and b _h are bias constants;

r_t的值越接近于1，则表示前一状态被写入的信息越多，r_t的越接近于0，则表示前一状态被写入的信息越少；r _t is used to control how much information from the previous state is written to the current candidate state

The closer the value of r _t is to 1, the more information is written in the previous state, and the closer the value of r _t is to 0, the less information is written in the previous state;

表示当前时刻下的候选状态信息，tanh(·)激活函数用于将值缩放到[-1,1]之间，即：

Represents the candidate state information at the current moment, and the tanh( ) activation function is used to scale the value between [-1, 1], namely:

The specific process of training the GRU prediction model in step 3) is as follows:

Suppose the mean square error loss function J(w) is used to evaluate the fitting degree of the GRU prediction model to the training data during training, and the expression of J(w) is:

为GRU预测模型预测的PUE值；Among them, N represents the number of training data, y _i represents the label of the i-th data, that is, the real PUE value,

PUE value predicted for the GRU prediction model;

Let the training optimizer be Adam, and the way to update the parameter w is:

First calculate the gradient g _t of the mean square error loss function J(w) to the parameter w, and then calculate the first-order moment m _t and the second-order moment v _t of the gradient g _t , where,

m _t =β ₁ ·m _t-1 +(1-β ₁ )·g _t

v _t =β ₂ ·v _t-1 +(1-β ₂ )·g _t ²

Among them, β ₁ and β ₂ are hyperparameters;

The first and second moments are then corrected by the following formulas;

Among them, β ₁ ^t and β ₂ ^t are the t powers of β ₁ and β ₂ , respectively.

The update process of parameter w is:

where α is the learning rate and ε is the hyperparameter.

The specific operation of step 4) is:

^The GRU prediction model is evaluated with the coefficient of determination R2, where the expression for the coefficient of determination R2 is ^:

为GRU预测模型的预测值，

为测试样本PUE的均值，决定系数R²表示GRU预测模型对真实数据的拟合程度，决定系数R²的值越接近1，则表示GRU预测模型的预测准确性越高。Among them, _yi is the real PUE value of the test sample,

is the predicted value of the GRU prediction model,

In order to test the mean value of the sample PUE, the coefficient of determination R ² represents the fit degree of the GRU prediction model to the real data. The closer the value of the coefficient of determination R ² is to 1, the higher the prediction accuracy of the GRU prediction model.

The present invention has the following beneficial technical effects:

The method for predicting the energy consumption efficiency value PUE of a data center based on the GRU neural network of the present invention selects the GRU prediction model as the prediction model during the specific operation. Because the GRU prediction model has the function of memorizing training information, it can make the attributes related to energy consumption. It is fully considered, such as the timing of temperature, humidity, etc., compared with the prior art, the correlation between PUE and energy consumption features is fully exploited in the training process, and the accuracy of predicting PUE is improved. It has been verified on the data generated by the simulator EnergyPlus, and compared with other traditional forecasting methods, and it has the best forecasting effect.

Description of drawings

1 is a schematic structural diagram of a network unit of a GRU;

FIG. 2 is a performance comparison diagram of the present invention and other prediction methods with correlation coefficient R ² as an evaluation index;

FIG. 3 is a fitting diagram of the predicted PUE value of the present invention and the real PUE.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in further detail:

1 , the method for predicting a data center energy consumption efficiency value PUE based on a GRU (GatedRecurrentUnit) neural network according to the present invention includes the following steps:

1) Collect attribute data related to data center energy consumption;

The specific operation of step 1) is: collecting the attribute data related to the energy consumption of the data center at a set time interval (10 min), wherein the collected attribute data related to the energy consumption of the data center includes attribute data related to IT equipment, and attribute data related to the environment. Attribute data, attribute data related to cooling system and attribute data related to infrastructure, among which attribute data related to IT equipment includes server load, UPS power load and thermal increase rate; environment related attribute data includes temperature, humidity and dew point temperature ; The attribute data related to the cooling system includes cooling rate and CRAC (Computer Room Air Condition) power; the attribute data related to infrastructure includes lighting and HVAC (Heating, Ventilation and Air Condition) power, set the data center energy consumption collected at time t Relevant attribute data X _t =(x ¹ , x ² , . . . , x ⁿ ).

2) Normalization and feature selection are performed on the attribute data collected in step 1);

The specific operations of step 2) are:

21) Normalize the collected attribute data X _t =(x ¹ ,x ² ,...,x ⁿ ) related to energy consumption of the data center to [0,1], and normalize the normalized data Denoted as X _i =(x ¹ ,x ² ,...,x ⁿ );

Assuming that the original data is Xi = (x ¹ , x ² ,..., ^{x n} ₎ , the normalized data attribute values are as follows:

Among them, x represents the original attribute value, x _min represents the minimum value of all data center attributes, x _max represents the maximum value of the attribute, x _normal represents the normalized attribute value, and its value is in [0,1] , denote the normalized data X _i as X _i =(x ¹ ,x ² ,...,x ⁿ );

22) Use a feature selection algorithm (RFECV (Recursive Feature Elimination with Cross-Validation)) to select m attribute data with the greatest correlation with the predicted PUE.

The specific operation of step 22) is:

Feature selection is performed using the RFECV method, and m attribute data X=(x ¹ , x ² , . . . , x ^m ) that are most relevant to the predicted PUE are selected.

3) Select the GRU prediction model as the prediction model, set the loss function J(w) and the optimization method optimizer, and then use the part of the data obtained by the feature selection in step 2) to train the GRU prediction model;

The specific operation of step 3) is:

In the existing prediction methods for the PUE value of data centers, machine learning methods such as simple ANN (Artificial Neural Network) and polynomial linear regression are used. The disadvantage of these methods is that they do not take into account the attributes related to data center energy consumption, such as: Temperature, humidity, etc. have time series characteristics. Therefore, in the present invention, it is proposed to use GRU (Gated Recurrent Unit) recurrent neural network as the prediction model of PUE.

The GRU recurrent neural network is used as the PUE prediction model, in which there are two gate units, reset gate r _t and update gate z _t in the GRU neuron structure in the GRU recurrent neural network, for To realize the training information generated by the training steps before memorizing or forgetting, the specific calculation process of the GRU unit is as follows:

z _t =σ(W _z ·[h _t-1 , x _t ]+b _z )

r _t =σ(W _r ·[h _t-1 , x _t ]+ _br )

Among them, z _t is used to control how much information h _t-1 of the previous state is written to the current state h _t , σ( ) is the sigmoid activation function that scales the value between (0, 1), namely:

t is the incoming parameter to be scaled, where the closer the value of σ( ) is to 0, the more writing, the closer the value of σ( ) is to 1, the less the writing is, where h _{t- 1} represents the state information output at time t-1, x _t represents the input at the current time, W _z , W _r and W _h are weight parameters, and b _z , _br and b _h are bias constants;

r_t的值越接近于1，则表示前一状态被写入的信息越多，r_t的越接近于0，则表示前一状态被写入的信息越少；r _t is used to control how much information from the previous state is written to the current candidate state

The closer the value of r _t is to 1, the more information is written in the previous state, and the closer the value of r _t is to 0, the less information is written in the previous state;

表示当前时刻下的候选状态信息，tanh(·)激活函数用于将值缩放到[-1,1]之间，即：

Represents the candidate state information at the current moment, and the tanh( ) activation function is used to scale the value between [-1, 1], namely:

The specific process of training the GRU prediction model in step 3) is as follows:

Suppose the mean square error loss function J(w) is used to evaluate the fitting degree of the GRU prediction model to the training data during training, and the expression of J(w) is:

为GRU预测模型预测的PUE值；Among them, N represents the number of training data, y _i represents the label of the i-th data, that is, the real PUE value,

PUE value predicted for the GRU prediction model;

Let the training optimizer be Adam, and the way to update the parameter w is:

First calculate the gradient g _t of the mean square error loss function J(w) to the parameter w, and then calculate the first-order moment m _t and the second-order moment v _t of the gradient g _t , where,

m _t =β ₁ ·m _t-1 +(1-β ₁ )·g _t

Among them, β ₁ and β ₂ are hyperparameters;

The first and second moments are then corrected by the following formulas;

Among them, β ₁ ^t and β ₂ ^t are the t powers of β ₁ and β ₂ , respectively.

The update process of parameter w is:

where α is the learning rate and ε is the hyperparameter.

4) Evaluate the trained GRU prediction model using the remaining data obtained from the feature selection in step 2), and then use the evaluated GRU prediction model to predict the energy consumption efficiency value PUE of the data center.

The specific operation of step 4) is:

The GRU prediction model is evaluated with the coefficient of determination R ² , where the expression for the coefficient of determination R ² is:

为GRU预测模型的预测值，

为测试样本PUE的均值，决定系数R²表示GRU预测模型对真实数据的拟合程度，决定系数R²的值越接近1，则表示GRU预测模型的预测准确性越高。Among them, _yi is the real PUE value of the test sample,

is the predicted value of the GRU prediction model,

In order to test the mean value of the sample PUE, the coefficient of determination R ² represents the fit degree of the GRU prediction model to the real data. The closer the value of the coefficient of determination R ² is to 1, the higher the prediction accuracy of the GRU prediction model.

As shown in Fig. 2, the prediction performance of the present invention on R ² is better than that of ANN, SVR and other prediction models. This is because GRU has the characteristics of memory, so that the time series characteristics of energy consumption attributes can be fully considered; GRU is better than The reason for LSTM is that GRU has fewer gated units, so there are fewer parameters, so the training time is less, the complexity is lower, and the prediction effect is better.

Claims (7)

1. A method for predicting a data center energy consumption efficiency PUE based on a GRU neural network is characterized by comprising the following steps:

1) collecting attribute data related to energy consumption of a data center;

2) normalizing the attribute data collected in the step 1) and selecting features;

3) selecting a GRU prediction model as a prediction model, setting a loss function J (w) and an optimization method optimizer, and then training the GRU prediction model by using part of data obtained by the feature selection in the step 2);

4) and (3) evaluating the trained GRU prediction model by utilizing the residual data obtained by the feature selection in the step 2), and predicting the energy consumption efficiency PUE of the data center by utilizing the evaluated GRU prediction model.

2. The method for predicting the data center energy consumption efficiency rate PUE based on the GRU neural network as claimed in claim 1, wherein the specific operation of the step 1) is as follows: collecting attribute data related to energy consumption of a data center according to a set time interval, wherein the collected attribute data related to the energy consumption of the data center comprises attribute data related to IT equipment, attribute data related to environment, attribute data related to a cooling system and attribute data related to infrastructure, and the attribute data related to the IT equipment comprises server load, electric load of a UPS and heat gain rate; the environment-related attribute data includes temperature, humidity, and dew point temperature; cooling system related attribute data includes cooling rate and CRAC power; the attribute data related to the infrastructure includes lighting and HVAC power, and the attribute data X related to the energy consumption of the data center collected at the time t_t＝(x¹，x²，...，xⁿ)。

3. The method for predicting the data center energy consumption efficiency rate PUE based on the GRU neural network as claimed in claim 2, wherein the specific operation of the step 2) is as follows:

21) collecting the obtained attribute data X related to the energy consumption of the data center_t＝(x¹，x²，...，xⁿ) Normalized to [0,1 ]]In between, the normalized data is recorded as X_i＝(x¹，x²，...，xⁿ)；

22) And selecting m attribute data with the maximum correlation with the predicted PUE by using a feature selection algorithm.

4. The method for predicting a data center energy consumption efficiency rate PUE based on a GRU neural network as claimed in claim 3, wherein the specific operation of step 22) is:

feature selection is performed by using an RFECV method, and m attribute data X with the maximum relevance to the predicted PUE are selected (X is X)¹，x²，...，x^m)。

5. The method for predicting PUE (energy consumption efficiency) of data center based on GRU (neural network) of claim 1, wherein in the step 3), the GRU recurrent neural network is used as a PUE prediction model, wherein a reset gate r exists in a GRU neuron structure in the GRU recurrent neural network_tAnd update gate z_tTwo gate control units, which are used for realizing the training information generated in the training step before memorizing or forgetting, wherein the specific calculation process of the GRU unit is as follows:

z_t＝σ(W_z·[h_t-1，x_t]+b_z)

r_t＝σ(W_r·[h_t-1，x_t]+b_r)

wherein z is_tFor controlling how much information h is in the previous state_t-1Is written to the current state h_tσ () is a sigmoid activation function that scales values between (0,1), i.e.:

t is the parameter to be scaled in, where a closer value of σ () to 0 means more writes and a closer value of σ () to 1 means less writes, where h_t-1Indicating the status information, x, output at time t-1_tRepresenting the input at the current time, W_z、W_rAnd W_hAs weight parameter, b_z、b_rAnd b_hIs a bias constant;

r_tfor controlling how much information from a previous state is written to a current candidate state

r_tThe closer to 1 the value of (d) is, the more information is written indicating the previous state, r_tThe closer to 0, the less information is written to the previous state;

representing candidate state information at the current time instant, tanh (-) activation function for scaling the value to [ -1,1]Namely:

6. the method for predicting the data center energy consumption efficiency PUE based on the GRU neural network as claimed in claim 5, wherein the specific process of training the GRU prediction model in the step 3) is as follows:

assuming that the mean square error loss function J (w) is used for evaluating the fitting degree of the GRU prediction model to the training data during training, the expression of J (w) is as follows:

where N denotes the number of training data, y_iThe tag representing the ith piece of data, i.e. the true PUE value,

PUE values predicted for GRU prediction models;

let the training optimizer Adam, and the way to update the parameter w is:

first, the gradient g of the mean square error loss function J (w) to the parameter w is calculated_tThen, the gradient g is calculated_tFirst moment m of_tAnd second moment v_tWherein, in the step (A),

m_t＝β₁·m_t-1+(1-β₁)·g_t

v_t＝β₂·v_t-1+(1-β₂)·g_t ²

wherein, β₁And β₂Is a hyper-parameter;

then, correcting the first moment and the second moment by the following formula;

wherein, β₁ ^tAnd β₂ ^tAre respectively β₁And β₂To the t power;

the updating process of the parameter w is as follows:

where α denotes the learning rate, and ε is a hyperparameter.

7. The method for predicting the data center energy consumption efficiency rate PUE based on the GRU neural network as claimed in claim 6, wherein the specific operation of the step 4) is as follows:

by determining the coefficient R²Evaluating the GRU prediction model, wherein a coefficient R is determined²The expression of (a) is:

wherein, y_iIn order to test the true PUE value of the sample,

for the predicted value of the GRU prediction model,

determining the coefficient R for the mean of the PUE of the test sample²Representing the fitting degree of the GRU prediction model to the real data, and determining a coefficient R²The closer to 1, the higher the prediction accuracy of the GRU prediction model.

Images