Method for optimizing refrigeration energy consumption of data center

Abstract

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

Description

Technical Field

[0001] This invention relates to the field of data center energy consumption control technology, and in particular to a method for optimizing data center cooling energy consumption. Background Technology

[0002] Data centers, as the core infrastructure of the information age, support the processing needs of various online services and massive amounts of data. With the rapid development of technologies such as the Internet, cloud computing, and big data, the scale of data centers is constantly expanding, and their energy consumption is growing exponentially, resulting in enormous energy consumption. IT (Information Technology) equipment and cooling equipment account for more than 80% of the total energy consumption. During IT equipment operation, heat is dissipated through hot aisles, heat exchange occurs through air conditioning return air, and then the cooled air is sent back to the IT equipment through cold aisles. During this cycle, the amount of heat dissipation determines the required cooling capacity, thus affecting cooling energy consumption. Cooling energy consumption can be directly controlled by adjusting parameters such as air temperature and airflow of cooling equipment (precision air conditioners). Exploring effective ways to save energy and reduce consumption in data centers is of great significance for promoting energy structure optimization and achieving green and sustainable development.

[0003] In the research on energy consumption control of data center cooling equipment, traditional methods such as PID control and rule-based control were initially used. These methods rely on fixed operating setpoints and conservative safety margins, lack dynamic adaptability in parameter adjustment, and are difficult to adapt to real-time fluctuations in IT load and the thermal inertia characteristics of the cooling system. Furthermore, they cannot coordinate the control requirements of mixed discrete and continuous actions, resulting in consistently low energy efficiency. In recent years, Model Predictive Control (MPC), Reinforcement Learning (RL), and some offline reinforcement learning methods have been gradually applied to this field.

[0004] While MPC has a certain global optimization capability, it is highly dependent on the environment model. When offline data is limited or the model accuracy is insufficient, it is prone to distributed out-of-control risks. At the same time, real-time computing costs are high.

[0005] Online RL methods require extensive real-time interaction with real cooling environments, have lengthy training cycles, and pose safety risks such as thermal violations and frequent equipment start-ups and shutdowns during the exploration phase, making them difficult to adapt to the high reliability requirements of data centers.

[0006] Offline RL methods are mostly designed for purely discrete or purely continuous action spaces, and are difficult to handle mixed action spaces; single-step decision-making tends to ignore the relationship between load mutation and system thermal inertia, and is difficult to balance the dual goals of energy consumption optimization and temperature control safety.

[0007] Analysis suggests that when the IT load in a data center suddenly increases, the aforementioned control methods are unable to accurately match the dynamic changes in cooling capacity and heat load, which can easily lead to energy waste from excessive cooling or operational risks of temperature runaway. Summary of the Invention

[0008] The purpose of this invention is to provide a method for optimizing data center cooling energy consumption. This method involves constructing an IT energy consumption prediction model and a decision model. The predicted IT energy consumption output from the prediction model and real-time collected sequence data of data center equipment operation are input into the decision model to obtain a decision value. The decision value is then used to optimize cooling energy consumption. This invention can reduce the energy consumption of cooling equipment. This invention is achieved through the following technical solutions.

[0009] This invention provides a method for optimizing cooling energy consumption in data centers, comprising:

[0010] Historical time-series data of data center equipment operation is acquired and preprocessed to obtain preprocessed historical time-series data, which is then used as training samples. The training samples include an IT energy consumption dataset for training an IT energy consumption prediction model and an offline dataset for training a decision model.

[0011] The IT energy consumption dataset and the offline dataset are input into the pre-built IT energy consumption prediction model and decision model, respectively, to obtain the trained IT energy consumption prediction model and decision model.

[0012] By inputting a portion of historical time-series data into the trained IT energy consumption prediction model, IT energy consumption prediction information can be obtained.

[0013] The IT energy consumption prediction information and real-time collected sequence data are input into the trained decision model to obtain decision values ​​corresponding to multiple future time points. The data center cooling energy consumption is then optimized based on these decision values.

[0014] Optionally, the preprocessed historical time series data (the previous historical moment) includes timestamps, historical IT energy consumption corresponding to the timestamps, total air conditioning energy consumption, return air temperature, inlet air temperature, supply air temperature, cold aisle temperature, hot aisle temperature, air volume, compressor start / stop status, historical IT energy consumption change amount, and historical IT energy consumption change rate.

[0015] The IT energy consumption dataset includes timestamp sequences and corresponding IT energy consumption data. The offline dataset includes historical IT energy consumption, total air conditioning energy consumption, return air temperature, inlet air temperature, supply air temperature, cold aisle temperature, hot aisle temperature, air volume, compressor start / stop status, historical IT energy consumption changes, and historical IT energy consumption change rates corresponding to the timestamps.

[0016] The aforementioned historical time series data (the first few historical moments) includes timestamp sequences and the corresponding historical IT energy consumption.

[0017] The real-time acquired sequence data includes real-time acquired IT energy consumption, return air temperature, inlet air temperature, cold aisle temperature, and hot aisle temperature.

[0018] Optionally, the IT energy consumption prediction model includes a sequentially connected input layer, time coding layer, input embedding layer, LSTM layer, local window attention layer, trend-aware attention layer, fully connected prediction layer, and output layer.

[0019] Optionally, the IT energy consumption prediction model specifically performs the following operations:

[0020] The input layer inputs a portion of historical time-series data to the time encoding layer. The time encoding layer transforms this data into features that the IT energy consumption prediction model can recognize, injecting time patterns to obtain historical time-series features that include these patterns. These features are then input to the input embedding layer. The input embedding layer unifies the dimensionality of the 11-dimensional original features, mapping them to a higher-dimensional space. This is then adapted to the subsequent LSTM network and attention layer, resulting in historical time-series features of a unified dimension, which are input to the LSTM layer. The LSTM layer captures long-range dependencies across time steps from these unified-dimensional historical time-series features, obtaining hidden-state sequences with consistent time-step lengths. These hidden-state sequences are then input to the local window attention layer. The local window attention layer performs local key dependency filtering on these hidden-state sequences with consistent time-step lengths, obtaining filtered local features that are input to the trend-aware attention layer. The trend-aware attention layer co-models the filtered local features with the global cycle, obtaining enhanced features that integrate local trend and global cycle information. These enhanced features are then input to the fully connected prediction layer. The fully connected prediction layer maps these enhanced features, which integrate local trend and global cycle information, into IT energy consumption prediction information and outputs it to the decision model through the output layer.

[0021] Optionally, the filtered local feature expression is as follows:

[0022] ,

[0023] In the formula, These are the local features after filtering. The hidden state sequence has a consistent time step length. Representation layer normalization, This indicates the attention of the bulls. Indicates a mask;

[0024] The expression for the enhanced feature that integrates local trend and global periodic information is as follows:

[0025] ,

[0026] In the formula, An enhanced feature that integrates local trend and global cycle information. It is a feedforward network. The output of multi-head attention is expressed as follows:

[0027] ,

[0028] In the formula, Q represents the query, K represents the key, and V represents the value. The expressions are as follows:

[0029] , This represents a 1×1 convolution with a kernel size of 5.

[0030] Optionally, the IT energy consumption prediction information includes predicted IT energy consumption values, changes in IT energy consumption, and the rate of change in IT energy consumption. The changes in IT energy consumption and the rate of change in IT energy consumption are calculated using the following formulas:

[0031] ,

[0032] ,

[0033] In the formula, This represents the change in IT energy consumption at the current moment. This represents the rate of change in IT energy consumption at the current moment. This represents the predicted IT energy consumption value at time t+1. This represents the actual IT energy consumption at the current time t.

[0034] Optionally, constructing the decision model includes extending the objective function of Conservative Q-Learning (CQL) to obtain an extended objective function, and then obtaining an offline reinforcement learning framework based on Hybrid Conservative Q-Learning (H-CQL) based on the extended objective function, thereby obtaining the decision model; wherein, the expression of the extended objective function is as follows:

[0035] ,

[0036] In the formula, For the expanded objective function, This represents the temporal difference loss, used to maintain consistency between the Q-value function and the Bellman objective. and Representing discrete actions With continuous action Conservative regularization term, and These are weighting coefficients used to balance the impact of different regularization terms on the overall objective.

[0037] Timing Differential Loss It is obtained by calculation using the following formula:

[0038] ,

[0039] In the formula, D is the offline dataset. For an enhanced state space, Samples based on offline dataset D Calculated expectation, The target return is calculated using the Bellman equation. The decomposed Q-function is obtained through the following formula:

[0040] ,

[0041] In the formula, Representing the enhanced state space Discrete Actions Its independent value In the enhanced state space and given discrete actions Next continuous action Conditional value, This represents the value of the discrete-continuous action interaction effect.

[0042] Optional, discrete actions conservative regularization term It is obtained by calculation using the following formula:

[0043] ,

[0044] In the formula, Indicates policy-based Discrete actions sampled from conditional probability distributions Calculated expectation, Represents the state sampled from the offline dataset D. Calculated expectation,

[0045] Continuous Actions conservative regularization term It is obtained by calculation using the following formula:

[0046] ,

[0047] In the formula, Represents continuous actions sampled from an offline dataset D. Calculated expectation, Represents the state sampled from the offline dataset D. and discrete actions Calculated expectation, This represents the state space in the given augmented state space. and given discrete actions Next continuous action The conditional value, of which This represents a continuous action sample obtained from the action distribution, where k is the number of samples, ranging from 1 to K, and K is the total number of samples.

[0048] Optionally, the multi-objective reward function expression of the decision model is as follows:

[0049] ,

[0050] In the formula, For the multi-objective reward function of the decision-making model, It is a discrete-continuous hybrid action. For energy consumption items, For comfort, This is a breach of contract item. , and All are weighting coefficients.

[0051] Optionally, the enhanced state space Including the basic state vector and predicted state vector The discrete-continuous hybrid action The expression is as follows:

[0052] ,

[0053] Discrete Actions Used to control the compressor's start / stop status, when When the value is 0, it indicates that the compressor has stopped. When the value is 1, it indicates that the compressor has started.

[0054] Continuous Actions Used to control supply air temperature and air volume The expression is as follows:

[0055] ,

[0056] Continuous Actions Physical constraints The expression is as follows:

[0057] .

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention constructs an IT energy consumption prediction model that accurately captures the periodicity and abrupt changes in IT energy consumption. By combining the decision value at the selected first moment, the calculated data center cooling energy consumption can be adjusted in advance based on future load trends, improving decision-making foresight and avoiding energy consumption fluctuations and frequent start-ups and shutdowns of cooling equipment caused by short-sighted decisions. An extended objective function is obtained by expanding the objective function of existing Conservative Q-Learning (CQL). ​​Based on this extended objective function, an offline reinforcement learning framework based on Hybrid Conservative Q-Learning (H-CQL) is derived, resulting in a decision model that effectively handles the discrete and continuous mixed action control requirements of data center cooling systems, alleviating the estimation distortion and training instability problems of traditional offline RL methods in mixed action spaces. Experimental verification shows that the proposed method can reduce energy consumption, and the constructed IT energy consumption prediction model can be trained based on offline historical data without real-time interaction with the real environment, avoiding the security risks of online exploration, and possessing high security and practicality. Attached Figure Description

[0060] Figure 1 The diagram shown is a schematic diagram of a data center cooling energy consumption optimization method in one embodiment of the present invention;

[0061] Figure 2 The diagram shown is a schematic diagram of the IT energy consumption prediction model structure in one embodiment of the present invention;

[0062] Figure 3 The diagram shown is a comparison of rewards for a data center cooling energy consumption optimization task in one embodiment of the present invention.

[0063] Figure 4 The diagram shown is a comparative illustration of data center IT energy consumption prediction in one embodiment of the present invention. Detailed Implementation

[0064] The following description, in conjunction with the accompanying drawings and specific embodiments, provides further details. In this description, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0065] Example 1

[0066] This embodiment introduces a method for optimizing data center cooling energy consumption, such as... Figure 1 As shown, it includes the following:

[0067] Historical time-series data of data center equipment operation is acquired and preprocessed to obtain preprocessed historical time-series data, which is then used as training samples. The training samples include an IT energy consumption dataset for training an IT energy consumption prediction model and an offline dataset for training a decision model.

[0068] The IT energy consumption dataset and the offline dataset are input into the pre-built IT energy consumption prediction model and decision model, respectively, to obtain the trained IT energy consumption prediction model and decision model.

[0069] By inputting a portion of historical time-series data into the trained IT energy consumption prediction model, IT energy consumption prediction information can be obtained.

[0070] The IT energy consumption prediction information and real-time collected sequence data are input into the trained decision model to obtain decision values ​​corresponding to multiple future time points. The data center cooling energy consumption is then optimized based on these decision values.

[0071] Example 2

[0072] Based on Example 1, this example describes the specific implementation process of a data center cooling energy consumption optimization method, which includes the following:

[0073] I. Data Preprocessing

[0074] In one specific embodiment of the present invention, the following steps are first taken: acquiring the timestamp sequence of a previous historical period, the corresponding historical IT energy consumption, total air conditioning energy consumption, return air temperature, inlet air temperature, supply air temperature, cold aisle temperature, hot aisle temperature, air volume, compressor start / stop status, historical IT energy consumption change, and historical IT energy consumption change rate. This data is then cleaned and filtered, normalized, standardized in the state and action space, and processed with reward signals to obtain preprocessed historical time series data. The preprocessed historical time series data is then divided into an IT energy consumption dataset for training an IT energy consumption prediction model and an offline dataset for training a decision-making model.

[0075] The IT energy consumption dataset includes timestamps and corresponding IT energy consumption. The offline dataset includes historical IT energy consumption, total air conditioning energy consumption, return air temperature, inlet air temperature, supply air temperature, cold aisle temperature, hot aisle temperature, air volume, compressor start / stop status, historical IT energy consumption changes, and historical IT energy consumption change rates.

[0076] Some historical time series data includes timestamp sequences and the corresponding historical IT energy consumption.

[0077] The real-time acquired sequence data includes real-time IT energy consumption, return air temperature, inlet air temperature, cold aisle temperature, and hot aisle temperature.

[0078] II. Constructing an IT Energy Consumption Prediction Model

[0079] In one specific embodiment of the present invention, the constructed IT energy consumption prediction model, such as Figure 2 As shown, it includes a sequentially connected input layer, temporal encoding layer, input embedding layer, LSTM layer, local window attention layer, trend-aware attention layer, fully connected prediction layer, and output layer.

[0080] In one specific embodiment of the present invention, the IT energy consumption prediction model performs the following operations:

[0081] The input layer feeds a portion of historical time-series data into the time encoding layer. The time encoding layer transforms this data into features that the IT energy consumption prediction model can recognize, injecting time patterns to obtain historical time-series features that include these patterns. These features are then fed into the input embedding layer. The input embedding layer unifies the dimensions (mapping the 11-dimensional original features to a higher-dimensional space to adapt to the subsequent LSTM network and attention layer), resulting in historical time-series features with unified dimensions. These features are then fed into the LSTM layer. The LSTM layer captures long-range dependencies across time steps from these unified-dimensional historical time-series features, obtaining hidden state sequences with consistent time step lengths. These hidden state sequences are then fed into the local window attention layer. The local window attention layer performs local key dependency filtering on these hidden state sequences with consistent time step lengths, obtaining filtered local features that are then fed into the trend-aware attention layer. The trend-aware attention layer co-models the filtered local features with the global cycle, obtaining enhanced features that integrate local trend and global cycle information. These enhanced features are then fed into the fully connected prediction layer. The fully connected prediction layer maps these enhanced features, which integrate local trend and global cycle information, into IT energy consumption prediction information and outputs it to the decision model through the output layer.

[0082] In one specific embodiment of the present invention, the filtered local feature expression is as follows:

[0083] ,

[0084] In the formula, These are the local features after filtering. The hidden state sequence has a consistent time step length. Representation layer normalization, This indicates the attention of the bulls. Indicates a mask;

[0085] The expression for the enhanced feature that integrates local trend and global cycle information is as follows:

[0086] ,

[0087] In the formula, An enhanced feature that integrates local trend and global cycle information. It is a feedforward network. The output of multi-head attention is expressed as follows:

[0088] ,

[0089] In the formula, Q represents the query, K represents the key, and V represents the value. The expressions are as follows:

[0090] , This represents a 1×1 convolution with a kernel size of 5.

[0091] In one specific embodiment of the present invention, the IT energy consumption prediction information includes the predicted IT energy consumption value, the change in IT energy consumption, and the rate of change in IT energy consumption. The change in IT energy consumption and the rate of change in IT energy consumption are calculated using the following formulas:

[0092] ,

[0093] ,

[0094] In the formula, This represents the change in IT energy consumption at the current moment. This represents the rate of change in IT energy consumption at the current moment. This represents the predicted IT energy consumption value at time t+1. This represents the actual IT energy consumption at the current time t.

[0095] III. Constructing a Decision Model

[0096] In one specific embodiment of the present invention, constructing a decision model includes extending the objective function of Conservative Q-Learning (CQL) to obtain an extended objective function, and obtaining an offline reinforcement learning framework based on Hybrid Conservative Q-Learning (H-CQL) based on the extended objective function, thereby obtaining the decision model; wherein, the expression of the extended objective function is as follows:

[0097] ,

[0098] In the formula, For the expanded objective function, This represents the temporal difference loss, used to maintain consistency between the Q-value function and the Bellman objective. and Representing discrete actions With continuous action Conservative regularization term, and These are weighting coefficients used to balance the impact of different regularization terms on the overall objective.

[0099] Timing Differential Loss It is obtained by calculation using the following formula:

[0100] ,

[0101] In the formula, D is the offline dataset. For an enhanced state space, Samples based on offline dataset D Calculated expectation, The target return is calculated using the Bellman equation. The decomposed Q-function is obtained through the following formula:

[0102] ,

[0103] In the formula, Representing the enhanced state space Discrete Actions Its independent value In the enhanced state space and given discrete actions Next continuous action Conditional value, This represents the value of the discrete-continuous action interaction effect.

[0104] In one specific embodiment of the present invention, discrete actions conservative regularization term It is obtained by calculation using the following formula:

[0105] ,

[0106] In the formula, Indicates policy-based Discrete actions sampled from conditional probability distributions Calculated expectation, Represents the state sampled from the offline dataset D. Calculated expectation,

[0107] Continuous Actions conservative regularization term It is obtained by calculation using the following formula:

[0108] ,

[0109] In the formula, Represents continuous actions sampled from an offline dataset D. Calculated expectation, Represents the state sampled from the offline dataset D. and discrete actions Calculated expectation, This represents the state space in the given augmented state space. and given discrete actions Next continuous action The conditional value, of which This represents a continuous action sample obtained from the action distribution, where k is the number of samples, ranging from 1 to K, and K is the total number of samples.

[0110] In one specific embodiment of the present invention, the multi-objective reward function expression of the decision model is as follows:

[0111] ,

[0112] In the formula, For the multi-objective reward function of the decision-making model, It is a discrete-continuous hybrid action. For energy consumption items, For comfort, This is a breach of contract item. , and All are weighting coefficients.

[0113] Energy consumption items The expression is as follows:

[0114] ,

[0115] In the formula, Total energy consumption for air conditioning This is the normalization constant.

[0116] Comfort item The expression is as follows:

[0117] ,

[0118] In the formula, The return air temperature represents the temperature value at time step t. The target temperature is the ideal temperature that is desired to be achieved. The maximum permissible temperature is the upper limit of the hard constraint. P is the penalty coefficient, used to adjust the degree of impact of temperature deviation on the reward.

[0119] Default The expression is as follows: ,

[0120] In the formula, to All are weighting coefficients. express,

[0121] In one specific embodiment of the present invention, an enhanced state space... Including the basic state vector and predicted state vector Basic state vector This includes IT energy consumption, return air temperature, hot aisle temperature, cold aisle temperature, and inlet air temperature. Predicted state vector. The output of the IT forecasting model, i.e., IT energy consumption forecast information, includes predicted IT energy consumption values, changes in IT energy consumption, and rates of change in IT energy consumption. Basic state vector. and predicted state vector The expressions are as follows:

[0122] ,

[0123] In the formula, Let be the actual IT energy consumption at time t. Let be the return air temperature at time t. Let be the temperature of the hot channel at time t. Let be the cold aisle temperature at time t. Let be the inlet air temperature at time t.

[0124] .

[0125] Discrete-continuous hybrid action The expression is as follows:

[0126] ,

[0127] Discrete Actions Used to control the compressor's start / stop status, when When the value is 0, it indicates that the compressor has stopped. When the value is 1, it indicates that the compressor has started.

[0128] Continuous Actions Used to control supply air temperature and air volume The expression is as follows:

[0129] ,

[0130] Continuous Actions Physical constraints The expression is as follows:

[0131] .

[0132] In one specific embodiment of the present invention, the constraints of the decision model are summarized as set C:

[0133] .

[0134] IV. Optimization of Data Center Cooling Energy Consumption

[0135] In one specific embodiment of the present invention, an enhanced state space... That is, the basic state vector and predicted state vector It is the input to the decision model, the basic state vector. Predict the state vector based on the real-time acquired sequence data. Corresponding IT energy consumption prediction information. Discrete actions. and continuous action It is the output of the decision model. Specifically, the output of the decision model is the compressor start / stop status and supply air temperature at multiple future points in time. and air volume .

[0136] This means that by using IT energy consumption forecasts and real-time collected sequence data, we can obtain the future load (IT equipment) fluctuations. Based on the future load fluctuations, we can obtain the actual compressor start / stop status, air supply temperature, and air supply volume. We can then adjust the compressor start / stop status, air supply temperature, and air supply volume of precision air conditioners (refrigeration equipment) in advance, thereby avoiding short-term energy waste caused by insufficient cooling capacity and long-term energy waste caused by short-sighted strategies, and thus achieving the effect of optimizing the energy consumption of refrigeration equipment.

[0137] In this embodiment, the compressor start / stop status and supply air temperature corresponding to the first moment output by the decision model are selected. and air volume As the decision value, the corresponding decision values ​​at other times are recalculated based on the input of the decision model at the next time moment, thus realizing a rolling optimization strategy. This method improves the foresight and stability of decision-making while maintaining computational feasibility, enabling the refrigeration equipment to respond in advance to changes in external load and IT equipment energy consumption, avoiding the delayed effects of energy consumption and temperature control.

[0138] V. Experiment

[0139] The validation dataset used in this embodiment employs energy consumption and sensor data from a data center in China to perform IT energy consumption prediction and cooling energy consumption optimization decisions. Table 1 compares the IT energy consumption prediction performance of the present invention's model with that of Autoformer, Informer, LSTM, and TCN on this dataset using evaluation metrics. By comparing the mean squared error, mean absolute error, and mean absolute percentage error of the other models, the present invention outperforms them.

[0140] Table 1. Comparison of running performance among different models

[0141] Figure 3 This diagram illustrates a comparison of the reward curves of the IT energy consumption prediction model constructed in this invention with those of TD3-BC, BCQ, BC, and the historical strategy (Historical) for the cooling energy consumption optimization decision task, under a selected dataset. The horizontal axis represents the time step, and the vertical axis represents the reward value. Figure 3 It can be seen that the IT energy consumption prediction model has a higher reward level and outperforms other models.

[0142] Figure 4 The diagram illustrates the IT energy consumption prediction model constructed in this invention, along with TCN, Autoformer, Informer, and LSTM, for predicting data center IT energy consumption under a selected dataset. The horizontal axis represents the time step, and the vertical axis represents the predicted IT energy consumption value. Figure 4 The True Values ​​in the table represent the actual IT energy consumption values. Figure 4 It can be seen that the prediction results of the IT energy consumption prediction model (LSTM+MA) constructed in this invention are closer to the actual values, indicating that the prediction results of this invention are more accurate.

[0143] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for optimizing cooling energy consumption in a data center, characterized in that, include: Historical time-series data of data center equipment operation is acquired and preprocessed to obtain preprocessed historical time-series data, which is then used as training samples. The training samples include an IT energy consumption dataset for training the IT energy consumption prediction model and an offline dataset for training the decision model. The IT energy consumption dataset and the offline dataset are input into the pre-built IT energy consumption prediction model and decision model, respectively, to obtain the trained IT energy consumption prediction model and decision model. By inputting a portion of historical time-series data into the trained IT energy consumption prediction model, IT energy consumption prediction information can be obtained. The IT energy consumption prediction information and real-time collected sequence data are input into the trained decision model to obtain decision values ​​corresponding to multiple future time points. The data center cooling energy consumption is then optimized based on these decision values.

2. The data center cooling energy consumption optimization method according to claim 1, characterized in that, The preprocessed historical time series data includes timestamp sequences, historical IT energy consumption corresponding to the timestamp sequences, total air conditioning energy consumption, return air temperature, inlet air temperature, supply air temperature, cold aisle temperature, hot aisle temperature, air volume, compressor start / stop status, historical IT energy consumption change amount, and historical IT energy consumption change rate. The IT energy consumption dataset includes timestamps and corresponding IT energy consumption. The offline dataset includes historical IT energy consumption, total air conditioning energy consumption, return air temperature, inlet air temperature, supply air temperature, cold aisle temperature, hot aisle temperature, air volume, compressor start / stop status, historical IT energy consumption change, and historical IT energy consumption change rate corresponding to the timestamps. The aforementioned historical time series data includes timestamp sequences of certain historical moments and the corresponding historical IT energy consumption. The real-time acquired sequence data includes real-time acquired IT energy consumption, return air temperature, inlet air temperature, cold aisle temperature, and hot aisle temperature.

3. The data center cooling energy consumption optimization decision method according to claim 1, characterized in that, The IT energy consumption prediction model includes a sequentially connected input layer, time coding layer, input embedding layer, LSTM layer, local window attention layer, trend-aware attention layer, fully connected prediction layer, and output layer.

4. The data center cooling energy consumption optimization method according to claim 3, characterized in that, The IT energy consumption prediction model specifically performs the following operations: The input layer inputs a portion of historical time-series data to the time encoding layer. The time encoding layer transforms this data into features that the IT energy consumption prediction model can recognize, injects time patterns, and obtains historical time-series features that include these patterns. These features are then input to the input embedding layer. The input embedding layer unifies the dimensions of these features to obtain historical time-series features with a unified dimension, which are then input to the LSTM layer. The LSTM layer captures long-range dependencies across time steps from these unified-dimensional historical time-series features, obtaining hidden-state sequences with consistent time-step lengths. These hidden-state sequences with consistent time-step lengths are then input to the local window attention layer. The local window attention layer performs local key dependency filtering on these hidden-state sequences to obtain filtered local features, which are then input to the trend-aware attention layer. The trend-aware attention layer co-models the filtered local features with the global cycle to obtain enhanced features that integrate local trend and global cycle information. These enhanced features are then input to the fully connected prediction layer. The fully connected prediction layer maps these enhanced features, which integrate local trend and global cycle information, to IT energy consumption prediction information and outputs them to the decision model through the output layer.

5. The data center cooling energy consumption optimization method according to claim 4, characterized in that, The filtered local feature expression is as follows: ， In the formula, These are the local features after filtering. The hidden state sequence has a consistent time step length. Representation layer normalization, This indicates the attention of the bulls. Indicates the mask; The expression for the enhanced feature that integrates local trend and global periodic information is as follows: ， In the formula, An enhanced feature that integrates local trend and global cycle information. It is a feedforward network. The output of multi-head attention is expressed as follows: ， In the formula, Q represents the query, K represents the key, and V represents the value. The expressions are as follows: , This represents a 1×1 convolution with a kernel size of 5.

6. The data center cooling energy consumption optimization method according to claim 4, characterized in that, The IT energy consumption prediction information includes predicted IT energy consumption values, changes in IT energy consumption, and the rate of change in IT energy consumption. The changes in IT energy consumption and the rate of change in IT energy consumption are calculated using the following formulas: ， ， In the formula, This represents the change in IT energy consumption at the current moment. This represents the rate of change in IT energy consumption at the current moment. This represents the predicted IT energy consumption value at time t+1. This represents the actual IT energy consumption at the current time t.

7. The data center cooling energy consumption optimization method according to claim 1, characterized in that, The decision model is constructed by extending the objective function of conservative Q-learning to obtain an extended objective function, and then deriving an offline reinforcement learning framework based on hybrid conservative Q-learning, thereby obtaining the decision model. The extended objective function is expressed as follows: ， In the formula, For the expanded objective function, Indicates the time-series differential loss. and Representing discrete actions With continuous action Conservative regularization term, and These are weighting coefficients used to balance the impact of different regularization terms on the overall objective. Timing Differential Loss It is obtained by calculation using the following formula: ， In the formula, D is the offline dataset. For an enhanced state space, Samples based on offline dataset D Calculated expectation, The target return is calculated using the Bellman equation. The decomposed Q-function is obtained through the following formula: ， In the formula, Representing the enhanced state space Discrete Actions Its independent value In the enhanced state space and given discrete actions Next continuous action Conditional value, This represents the value of the discrete-continuous action interaction effect.

8. The data center cooling energy consumption optimization method according to claim 7, characterized in that, Discrete Actions conservative regularization term It is obtained by calculation using the following formula: ， In the formula, Indicates policy-based Discrete actions sampled from conditional probability distributions Calculated expectation, Represents the state sampled from the offline dataset D. Calculated expectation, Continuous Actions conservative regularization term It is obtained by calculation using the following formula: ， In the formula, Represents continuous actions sampled from an offline dataset D. Calculated expectation, Represents the state sampled from the offline dataset D. and discrete actions Calculated expectation, This represents the state space in the given augmented state space. and given discrete actions Next continuous action The conditional value, of which This represents a continuous action sample obtained from the action distribution, where k is the number of samples, ranging from 1 to K, and K is the total number of samples.

9. The data center cooling energy consumption optimization method according to claim 8, characterized in that, The multi-objective reward function expression of the decision model is as follows: ， In the formula, For the multi-objective reward function of the decision-making model, It is a discrete-continuous hybrid action. For energy consumption items, For comfort, This is a breach of contract item. , and All are weighting coefficients.

10. The data center cooling energy consumption optimization method according to claim 9, characterized in that, The enhanced state space Including the basic state vector and predicted state vector The discrete-continuous hybrid action The expression is as follows: ， Discrete Actions Used to control the compressor's start / stop status, when When the value is 0, it indicates that the compressor has stopped. When the value is 1, it indicates that the compressor has started. Continuous Actions Used to control supply air temperature and air volume The expression is as follows: ， Continuous Actions Physical constraints The expression is as follows: 。

Images