CN-121996956-A - Data center refrigeration energy consumption optimization method

CN121996956ACN 121996956 ACN121996956 ACN 121996956ACN-121996956-A

Abstract

The invention discloses a data center refrigeration energy consumption optimization method, which belongs to the technical field of data center energy consumption control and comprises the steps of obtaining historical time series data of data center machine room equipment operation, preprocessing the historical time series data to obtain preprocessed historical time series data, taking the preprocessed historical time series data as training samples, respectively inputting the training samples into a pre-constructed IT energy consumption prediction model and a pre-constructed decision model to obtain a trained IT energy consumption prediction model and a trained decision model, inputting part of historical time series data into the trained IT energy consumption prediction model to obtain IT energy consumption prediction information, inputting the IT energy consumption prediction information and sequence data acquired in real time into the trained decision model to obtain decision values corresponding to a plurality of moments in the future, and optimizing the refrigeration energy consumption of the data center through the decision values, so that the energy consumption of refrigeration equipment can be reduced.

Inventors

LI LIJUAN
LIU HEYU
JIAO WENHUA
CHEN JIANXUN

Assignees

南京工业大学

Dates

Publication Date: 20260508
Application Date: 20260114

Claims (10)

1. The data center refrigeration energy consumption optimizing method is characterized by comprising the following steps: Acquiring historical time sequence data of the operation of the data center machine room equipment, preprocessing the historical time sequence data to obtain preprocessed historical time sequence data, and taking the preprocessed historical time sequence data as a training sample, wherein the training sample comprises an IT energy consumption data set for training an IT energy consumption prediction model and an offline data set for training a decision model; Respectively inputting the IT energy consumption data set and the offline data set into a pre-constructed IT energy consumption prediction model and a decision model to obtain a trained IT energy consumption prediction model and a trained decision model; inputting part of historical time series data into the trained IT energy consumption prediction model to obtain IT energy consumption prediction information, And (3) inputting the IT energy consumption prediction information and the sequence data acquired in real time into a trained decision model to obtain decision values corresponding to a plurality of moments in the future, and optimizing the refrigeration energy consumption of the data center through the decision values.
2. The method for optimizing refrigeration energy consumption of a data center according to claim 1, wherein the preprocessed historical time series data comprises a time stamp sequence, historical IT energy consumption corresponding to the time stamp sequence, total air conditioner energy consumption, return air temperature, inlet air temperature, supply air temperature, cold channel temperature, hot channel temperature, supply air volume, start-stop state of a compressor, historical IT energy consumption change amount and historical IT energy consumption change rate, The IT energy consumption data set comprises a time stamp and IT energy consumption corresponding to the time stamp, the offline data set comprises historical IT energy consumption, total air conditioner energy consumption, return air temperature, inlet air temperature, supply air temperature, cold channel temperature, hot channel temperature, supply air quantity, compressor start-stop state, historical IT energy consumption change quantity and historical IT energy consumption change rate corresponding to the time stamp, The partial history time series data comprises a time stamp series of partial history time and a history IT energy consumption corresponding to the time stamp series, The sequence data acquired in real time comprises IT energy consumption, return air temperature, inlet air temperature, cold channel temperature and hot channel temperature acquired in real time.
3. The data center refrigeration energy consumption optimization decision-making method according to claim 1, wherein the IT energy consumption prediction model comprises an input layer, a time coding layer, an input embedded layer, an LSTM layer, a local window attention layer, a trend perception attention layer, a full connection prediction layer and an output layer which are connected in sequence.
4. The data center refrigeration energy consumption optimization method of claim 3, wherein the IT energy consumption prediction model specifically performs the following operations: The input layer inputs partial historical time series data to the time coding layer, the time coding layer converts the partial historical time series data into characteristics which can be identified by the IT energy consumption prediction model, the time law is injected, the historical time series characteristics comprising the time law are obtained and input to the input embedded layer, the input embedded layer unifies the historical time series characteristics with unified dimension and inputs the historical time series characteristics to the LSTM layer, the LSTM layer captures the long-range dependence of the unified dimension on the historical time series characteristics with consistent time steps to obtain hidden state sequences with consistent time step length and inputs the hidden state sequences to the local window attention layer, the local window attention layer performs local critical dependence screening on the hidden state sequences with consistent time step length to obtain screened local characteristics and inputs the screened local characteristics to the trend perception attention layer, the trend perception attention layer models the screened local characteristics and the global period in a synergic mode to obtain enhanced characteristics fused local trend and global period information and inputs the enhanced characteristics into the fully-connected prediction layer, and the fully-connected prediction layer maps the enhanced characteristics fused local trend and global period information to IT energy consumption prediction information and outputs the IT energy consumption prediction information to the decision model through the output layer.
5. The data center refrigeration energy consumption optimization method according to claim 4, wherein the screened local feature expression is as follows: , in the formula, In order to screen out the local features after the selection, For a sequence of hidden states of consistent time step size, The representation layer is normalized and, Representing the attention of the multiple heads, A representation mask; the expression of the enhancement features fusing the local trend and the global period information is as follows: , in the formula, To fuse the local trend with the enhancement features of the global period information, In order to be a feed-forward network, The output representing the multi-headed attention is expressed as follows: , Wherein Q represents a query, K represents a key, V represents a value, and the expressions are as follows: , A 1 x1 convolution with a convolution kernel size of 5 is represented.
6. The method for optimizing refrigeration energy consumption of data center according to claim 4, wherein the IT energy consumption prediction information includes an IT energy consumption prediction value, an IT energy consumption variation amount and an IT energy consumption variation rate, and the IT energy consumption variation amount and the IT energy consumption variation rate are calculated by the following formulas: , , in the formula, Represents the change amount of IT energy consumption at the current moment, Indicating the rate of change of IT energy consumption at the current time, Represents the IT energy consumption predicted value at time t +1, The actual value of IT energy consumption at the current time t is indicated.
7. The method for optimizing refrigeration energy consumption of a data center according to claim 1, wherein constructing a decision model comprises expanding a conservative Q learning objective function to obtain an expanded objective function, and obtaining an offline reinforcement learning framework based on hybrid conservative Q learning based on the expanded objective function to obtain the decision model, wherein the expression of the expanded objective function is as follows: , in the formula, In order to expand the objective function after the expansion, Representing the loss of time-series differential, And (3) with Respectively represent discrete actions And continuous action Is used to determine the term for the conservative regularization of (1), And (3) with For the weighting coefficients, for balancing the influence of different regularization terms in the overall objective, Differential loss of time sequence Obtained by calculation by the following formula: , Where D is the offline data set, In order to enhance the state space of the device, For samples based on sampling in the offline dataset D The desire for the calculation is that, For the return objective calculated by the bellman equation, For the decomposition Q function, it is obtained by the following formula: , in the formula, Representing an enhanced state space Discrete action Is added to the product of the present invention, Represented in an enhanced state space Given discrete actions Lower continuous motion Is used to determine the value of the condition, Representing the value of the discrete action-continuous action interaction effect.
8. The data center refrigeration energy consumption optimization method of claim 7, wherein the discrete actions Is a conservative regularization term of (2) Obtained by calculation by the following formula: , in the formula, Representation based on policy Discrete actions sampled from conditional probability distributions The desire for the calculation is that, Representing states based on sampling in offline dataset D The desire for the calculation is that, Continuous motion Is a conservative regularization term of (2) Obtained by calculation by the following formula: , in the formula, Representing continuous actions based on sampling in offline dataset D The desire for the calculation is that, Representing states based on sampling in offline dataset D Discrete actions The desire for the calculation is that, Expressed in a given enhanced state space Given discrete actions Lower continuous motion Of (3), wherein And representing continuous action samples obtained by sampling in the action distribution, wherein K is the sampling number, the value range is 1-K, and K is the total sampling number.
9. The data center refrigeration energy consumption optimization method of claim 8, wherein the multi-objective rewards function expression of the decision model is as follows: , in the formula, For a multi-objective rewards function of the decision model, In the case of a discrete-continuous mixing action, In order to be an energy consumption term, In order to be a comfort level item, In the event of a breach of the contract, 、 And Are weight coefficients.
10. The data center refrigeration energy consumption optimization method of claim 9, wherein the enhanced state space Comprising a basic state vector And predicting a state vector The discrete-continuous mixing action The expression is as follows: , Discrete actions For controlling the start-stop state of the compressor when When 0 is taken, the compressor is stopped, when When 1 is taken, the start of the compressor is indicated; Continuous motion For controlling the supply air temperature And air supply amount The expression is as follows: , Continuous motion Subject to physical constraints The expression is as follows: 。

Description

Data center refrigeration energy consumption optimization method Technical Field The invention relates to the technical field of data center energy consumption control, in particular to a data center refrigeration energy consumption optimization method. Background The data center is used as a core infrastructure of the information age, and supports the processing requirements of various online services and mass data. With the rapid development of technologies such as internet, cloud computing and big data, the scale of a data center is continuously enlarged, and the energy consumption of the data center is exponentially increased, so that huge energy consumption is caused. Wherein IT (Information Technology) equipment and refrigeration equipment account for more than 80% of the total energy consumption. When the IT equipment operates, heat is radiated through the heat channel, heat exchange is carried out through air conditioning return air, and cooled air is sent into the IT equipment through the cold channel. During this cycle, the heat dissipation determines the amount of refrigeration required, thereby affecting the refrigeration energy consumption. The refrigeration energy consumption can be directly controlled by controlling the control quantity such as the air temperature, the air quantity and the like of refrigeration equipment (precision air conditioner), so that an effective way of energy conservation and consumption reduction of a data center is searched, and the method has important significance in promoting the optimization of an energy structure and realizing green sustainable development. In the research of energy consumption control of refrigeration equipment in a data center, traditional methods such as PID control and rule-based control are initially adopted, the methods depend on a fixed operation set point and a conservative safety margin, the parameter adjustment lacks dynamic adaptability, the real-time fluctuation of an IT load and the thermal inertia characteristic of a refrigeration system are difficult to adapt, and the mixed control requirements of discrete action and continuous action cannot be cooperatively processed, so that the energy utilization efficiency is in a low level for a long time. In recent years, model Predictive Control (MPC), reinforcement Learning (RL), and partially offline reinforcement learning methods are increasingly applied to this field: Although the MPC has certain global optimization capability, the MPC has strong dependence on an environment model, the MPC is easy to generate control risks outside distribution when offline data is limited or model accuracy is insufficient, and meanwhile, the real-time calculation cost is high; The online RL method needs to interact with a real refrigeration environment in a large amount in real time, has long training period, has safety risks of thermal violations and frequent start and stop of equipment in an exploration stage, and is difficult to adapt to the high-reliability operation requirement of the data center; the off-line RL method is mainly aimed at the problems that the design of a pure discrete or pure continuous action space is difficult to process a mixed action space, the association of the load mutation and the thermal inertia of a system is easy to be ignored in single-step decision, and the double targets of energy consumption optimization and temperature control safety are difficult to balance. According to analysis, when the IT load of the data center suddenly rises, the control method is difficult to accurately match the dynamic changes of the refrigerating capacity and the heat load, and the excessive refrigeration energy consumption waste or the running risk of temperature runaway are easily caused. Disclosure of Invention The invention aims to provide a data center refrigeration energy consumption optimization method, which is characterized in that an IT energy consumption prediction model and a decision model are constructed, the IT energy consumption prediction value output by the IT energy consumption prediction model and the sequence data which is acquired in real time and is operated by data center machine room equipment are input into the decision model to obtain a decision value, and the refrigeration energy consumption is optimized through the decision value. The invention is realized by the following technical scheme. The invention provides a data center refrigeration energy consumption optimization method, which comprises the following steps: Acquiring historical time sequence data of the operation of the data center machine room equipment, preprocessing the historical time sequence data to obtain preprocessed historical time sequence data, and taking the preprocessed historical time sequence data as a training sample, wherein the training sample comprises an IT energy consumption data set for training an IT energy consumption prediction model and an offline data set for training a de