Data center air conditioner energy-saving control method based on joint prediction and reinforcement learning
By combining the XGBoost temperature prediction model and the DQN control strategy, the system achieves proactive regulation of the data center air conditioning system, solving the problems of high energy consumption and lag in regulation, significantly reducing energy consumption and improving safety.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NATIONAL COMPUTER NETWORK & INFORMATION SECURITY MANAGEMENT CENTER YUNNAN BRANCH
- Filing Date
- 2026-03-04
- Publication Date
- 2026-05-08
AI Technical Summary
Data center air conditioning systems are energy-intensive and have slow adjustment. Traditional control methods lack predictability, and conventional reinforcement learning algorithms have slow response and cannot effectively cope with sudden load changes, resulting in wasted cooling capacity and safety hazards.
A temperature prediction model based on the XGBoost algorithm is combined with a deep reinforcement learning (DQN) control strategy. By predicting future temperature changes and embedding them into the reinforcement learning state space, the air conditioning parameters are optimized to achieve forward-looking decision-making.
Energy consumption was reduced by 12.3%, the duration of temperature exceeding the standard was reduced by 81%, and the PUE was optimized to 1.21, improving the energy efficiency and safety of the air conditioning system.
Smart Images
Figure CN121995767A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data center air conditioning and cooling control technology, specifically relating to a data center air conditioning energy-saving control method based on joint prediction and reinforcement learning. Background Technology
[0002] Uneven IT load within data center server rooms leads to uneven heat and cold distribution, and airflow resistance causes mutual energy consumption between air conditioners. IT load fluctuates and varies between day and night, while outdoor temperatures also fluctuate seasonally. Simple adjustments, while maintaining the temperature within a normal range, cannot anticipate load changes and implement overall control strategies, preventing air conditioners from operating optimally. The uneven temperature and humidity within the server room, coupled with mutual interference between air conditioners, not only results in significant waste of cooling capacity but also causes frequent start-stop cycles under fluctuating conditions, posing safety hazards. Furthermore, data center air conditioning systems account for up to 40% of energy consumption. Traditional control methods, such as PDI or threshold rules, lack predictability, forcing precision air conditioners to operate according to simple control logic, often leading to over-cooling or under-cooling. Conventional reinforcement learning (RL) algorithms only respond to historical / current states and cannot predict sudden load changes or thermal inertia effects, exhibiting response lag and insufficient predictability. Conventional multi-agent RL also suffers from high complexity and difficulty in convergence. Summary of the Invention
[0003] Purpose of the invention: To address the technical problems and defects existing in the prior art, this invention provides a data center air conditioning energy-saving control method based on joint prediction and reinforcement learning. The model control strategy of this invention has look-ahead decision-making capability, which can effectively reduce energy consumption by 12.3%, reduce the duration of temperature exceeding the standard by 81%, and optimize PUE to 1.21. It solves the problems of adjustment lag or overcooling in conventional control methods. This invention uses the predicted value as an extended state of part of the observable MDP to solve the short-sightedness problem of traditional RL.
[0004] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A data center air conditioning energy-saving control method based on joint prediction and reinforcement learning includes the following steps: S1, Collect historical data: Collect environmental data of the data center computer room during historical periods, including: time information, location temperature, outdoor temperature, air conditioning parameters and dynamic indicators; S2, Prediction Model Training: A data center temperature prediction model is built based on the XGBoost algorithm, and the historical data of the data center collected in S1 is used as training samples to complete the model training and optimization, and obtain the trained data center temperature prediction model. S3 acquires the current environmental data of the data center and, based on the model trained in S2, uses the current environmental data as input features to output the temperature prediction result through the optimized temperature prediction model. S4. Embed the temperature prediction results obtained in step S3 into the environmental data of the data center to obtain the data center state S. By incorporating the prediction information into the state space, DQN can make forward-looking decisions. S5. Construct a DQN reinforcement learning model and a reward function R. Take the data center state S and the control quantity A of the air conditioning unit action execution as input data, and output the updated control quantity A' of the air conditioning unit action execution. The updated control quantity A' of the air conditioning unit action execution is applied to the data center to obtain the updated data center PUE value and reward value R. S6. Repeat steps S4 and S5 to perform multiple rounds of iterative training and optimization on the DQN reinforcement learning model. When the data center PUE value and the reward function feedback value R tend to stabilize, the trained DQN reinforcement learning model is obtained. S7. Based on the DQN reinforcement learning model trained in step S6, the data center state S, which incorporates the temperature prediction results from the XGBoost algorithm, is used as input data to complete the output of the data center air conditioning energy-saving control quantity A.
[0005] Furthermore, in step S1, the air conditioning parameters include supply air temperature, return air temperature, fan speed, fan frequency, and compressor frequency; the dynamic indicators include temperature change rate and IT load power change; in step S2, the temperature prediction result is the maximum temperature of the data center in the next 5 minutes. .
[0006] Furthermore, in step S2, the XGBoost algorithm uses a serial CART regression tree to progressively optimize the loss function to predict the ambient temperature of the data center. The constructed loss function L is as follows: , In the formula, This represents the true temperature value of the i-th sample. This represents the predicted temperature value for the i-th sample. The tree complexity regularization parameter is used to control splitting, and its value is 0.1. This represents the total number of CART regression trees. This represents the L2 regularization coefficient, with a value of 0.1. The L2 norm represents the weight of the leaf node.
[0007] Furthermore, the reward function R described in step S4 is designed as follows: , In the formula, T represents the maximum temperature of the data center in the next 5 minutes. threshold The temperature alarm threshold is set to 25℃, E represents the data center energy consumption, α represents the weighting coefficient for the balanced temperature control progress, and β represents the weighting coefficient for energy consumption optimization.
[0008] Furthermore, in step S1, the sample data undergoes dimensionality reduction processing according to the following steps: S1.1, Standardize the environmental data to obtain a standardized feature vector. The standardization preprocessing includes missing value processing, outlier removal, and normalization processing. S1.2, Based on feature engineering and prior knowledge of engineering structures, optimize and recommend indoor temperature points closely related to outdoor temperature, air conditioning parameters and dynamic indicators through expert identification; (This step is used to screen temperature points sensitive to changes in outdoor temperature, temperature zones (multiple temperature points) with control association with specific air conditioning units, and high-risk areas for temperature control imbalance). S1.3, Based on Principal Component Analysis (PCA), eigenvalue vectors are obtained for each temperature point according to their covariance, and then sorted according to the eigenvalues. The following formula is used to ensure that the cumulative covariance is greater than a preset threshold. Identify the minimum required features in time. , In the formula, x j Let j represent the j-th feature, k represent the total number of features, and f represent the cumulative covariance ratio function. A larger function value indicates a better fit. This indicates a preset threshold, ranging from 85% to 95%.
[0009] Furthermore, the iterative training and optimization process of the DQN reinforcement learning model described in step S5 is as follows: S5.1, Initialize the DQN reinforcement learning model policy and target network; S5.2, At the current time step, based on the current state, use the ε algorithm to determine the control quantity of the action to be executed in the action network; S5.3 After executing action A on the computer room air conditioner according to the control quantity, update the air conditioning unit parameter settings and predict the temperature data for the next five minutes through the XGBoost temperature prediction model, and obtain the maximum temperature value T in the predicted temperature. S5.4 Construct a new data center state S, calculate the reward value of the action based on the reward function R, and feed the reward value of the action back to the DQN reinforcement learning model to update the model policy. S5.5, Repeat steps S5.2-S5.4 until the reward value of the action tends to stabilize, and complete the iterative training and optimization of the model. Furthermore, an experience replay pool is constructed, which stores the data center state S, the action A, the reward value R, and the new data center state S' obtained in step S5.5 for each time step into the experience replay pool. Every certain number of time steps, samples are taken from the experience replay pool, and the model weights are updated using DQN to complete the periodic update of the target network.
[0010] Furthermore, in step S6, the iterative training and optimization of the DQN reinforcement learning model ends when the following condition is met: , In the formula, This represents the reward value in the j-th iteration. This represents the reward value in the (j+n)th iteration. It is the constraint coefficient, with a value range of 0.01-0.1.
[0011] Furthermore, by performing Pearson correlation analysis on the input features of the XGBoost temperature prediction model, variables strongly correlated with future temperature prediction were selected. These strongly correlated variables include the air conditioning supply temperature setpoint, the location temperature, the average location temperature over a 20-minute period, and the temperature change rate.
[0012] Furthermore, during iterative training of the DQN reinforcement learning model, The action network uses a multilayer perceptron (MLP) with a structure consisting of a 64-node input layer and a 32-node hidden layer, both of which use the ReLU activation function. The output layer dimension is the action space size |A|. The target network is a copy of the action network and is trained stably through a delayed update mechanism with a delayed update coefficient τ=0.01 (i.e., synchronizing the policy network weights every 100 steps).
[0013] Beneficial Effects: Compared with existing technologies, this paper proposes a joint optimization framework integrating XGBoost temperature prediction and deep reinforcement learning (DQN). This framework uses an XGBoost model to accurately predict the data center temperature for the next 5 minutes (RMSE < 0.5℃, accuracy > 92%), and embeds the predicted values into the reinforcement learning state space, enabling the control strategy to have forward-looking decision-making capabilities. Furthermore, expert-guided PCA feature reduction and the air conditioning-temperature zone correlation matrix are introduced to improve the model's convergence speed and generalization performance. Real-world testing in a large data center shows that compared to PID control, this method reduces energy consumption by 12.3%, decreases the duration of temperature exceedance by 81%, and optimizes PUE to 1.21. This invention provides a reliable solution for data center energy saving. Attached Figure Description
[0014] Figure 1This is a logical flowchart of the data center air conditioning energy-saving control method based on joint prediction and reinforcement learning described in this invention.
[0015] Figure 2 This is a trend graph showing the changes in highly relevant input features according to embodiments of the present invention.
[0016] Figure 3 The results show the hourly air conditioning power consumption comparison in the embodiments of the present invention using PID and the control mode of the present invention, respectively.
[0017] Figure 4 This is a summary comparison of the power consumption of the computer room air conditioning in April 2025 using PID and other strategies, and the control mode of this invention. Detailed Implementation
[0018] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.
[0019] This invention presents a data center air conditioning energy-saving control method based on joint prediction and reinforcement learning. It is an intelligent control method that combines an XGBoost prediction model with a reinforcement learning control strategy. First, the invention utilizes an XGBoost temperature prediction model to accurately predict temperature changes over the next few minutes. The prediction results are then input into a reinforcement learning controller as part of the environmental state, enabling dynamic adjustment of air conditioning parameters. By introducing forward-looking prediction information, the reinforcement learning strategy possesses stronger predictive and reactive capabilities, effectively reducing energy consumption while ensuring temperature stability. This method effectively improves the safety and energy efficiency of the control strategy while maintaining system response sensitivity.
[0020] like Figure 1 As shown, the present invention specifically includes the following: I. XGBoost Temperature Prediction Model Algorithm
[0021] XGBoost, the look-ahead module of this invention, is an ensemble learning algorithm based on gradient trees. It improves overall prediction performance by sequentially constructing multiple weak classifiers (i.e., CART regression trees) and progressively optimizing the loss function. The XGBoost algorithm model used in this invention takes data center environmental data as input, including air conditioning supply air temperature setpoints, air conditioning return air temperature setpoints, point temperatures (data collected from sensors at various points in the server room), fan speeds, outdoor temperatures, historical temperature statistics (including temperature statistics for the 10 / 20 / 30 minutes prior to the prediction time), temperature change rate, and IT load power changes. The output prediction target is the server room temperature data for the next 5 minutes, used to assess the potential high-temperature risk in the server room caused by the current adjustment strategy. The XGBoost temperature prediction model is constructed as follows: Input feature X, , It refers to the air conditioner's air supply temperature. Air conditioner return air temperature It is the point temperature. It is the fan speed. It's the outdoor temperature. This represents the average temperature over the previous 20 minutes. Indicates the rate of temperature change. This indicates changes in IT load power.
[0022] Target output , ,in: It is the output function of the model, which obtains the final predicted value by summing the predictions of K trees. , The k-th CART regression tree (weak classifier) is the base learner in the model. This represents the function space of the CART regression tree, which is the set of all possible trees.
[0023] And construct the following loss function L, , In the formula, This represents the true temperature value of the i-th sample. This represents the predicted temperature value for the i-th sample. This represents the tree complexity regularization parameter (default 0.1, controls splitting). This represents the total number of CART regression trees. This represents the L2 regularization coefficient (default 1.0). The L2 norm represents the weight of the leaf node.
[0024] Specifically, the XGBoost regression model is trained on historical data from the data center. By predicting changes in factors such as air conditioning parameters, IT load, and outdoor temperature, it can provide corresponding air conditioning operating parameters in advance, preventing overheating and ensuring the data center's safety. The maximum temperature value predicted by XGBoost 5 minutes later is used to construct the environment state for the reinforcement learning algorithm.
[0025] II. Reinforcement Learning Model DQN
[0026] This invention employs a deep network, DQN, which is based on a Markov decision process and consists of a state space, an action space, transition probabilities, and a reward function. At each time step, an action is selected and executed. Subsequently, based on environmental feedback, a reward is obtained for the action, and a new state is entered. The policy is learned and optimized by maximizing the cumulative reward.
[0027] This invention, DQN, utilizes a neural network to approximate the action-value function, enabling it to handle high-dimensional state space information. In a temperature control scenario, the state includes the current air conditioning settings (such as supply air temperature setting, return air temperature setting, fan speed, etc.), the temperature point in the server room, the outdoor temperature, and the temperature prediction result. The actions are combinations of adjusting several air conditioning supply and return air temperature settings and turning the units on and off. The reward function comprehensively considers the degree of temperature exceeding the limit and the energy consumption of the server room. Compared to traditional methods that rely on current state-based decision-making, which lacks awareness of future trends and leads to lag and instability in algorithm strategies, this invention embeds XGBoost prediction results into the state construction of reinforcement learning, enabling the policy network to have a forward-looking judgment ability on future temperature trends, thereby achieving prediction-driven policy optimization. Detailed explanations are as follows. (1) Construction of the deep network model DQN: State definition: , This indicates the current operating status of the air conditioning (A / C) system.
[0028] Action definition: In the formula, Indicates the temperature adjustment range. Indicates the fan speed activation level. This indicates the start / stop status of IT.
[0029] Reward function R: Where T is the maximum predicted temperature. The temperature alarm threshold is 25℃, E is the energy consumption corresponding to the change in air conditioner fan speed, and α and β are the weighting coefficients for balancing temperature control accuracy and energy consumption optimization, respectively. This reward function will impose a secondary penalty on temperature exceeding the limit to strengthen the suppression of the risk of temperature exceeding the limit, so as to encourage the strategy to minimize energy consumption while avoiding temperature exceeding the limit.
[0030] (2) Training process of deep network model DQN
[0031] In each training round, the DQN model selects the action to be executed based on the current state. It then uses the XGBoost temperature prediction model to predict the temperature change after executing the action, calculates the reward for the action and the next state, and updates the policy network weights using DQN. During training, an experience replay and a greedy ε-policy are used to balance exploration and utilization. The specific training process is as follows: S5.1, Initialize the policy network and the target network; S5.2, at each time step, select the action to be executed in the action network based on the current state using the greedy ε algorithm; S5.3 After executing action A, update the air conditioning parameter settings and predict the temperature data for the next five minutes using the XGBoost temperature prediction model, then obtain the maximum temperature value from the predicted temperature. ; S5.4 Construct a new state S and calculate the reward value of the action according to the reward function R; feed the reward value of the action into the DQN reinforcement learning model to update the model policy; S5.5, store (state, action, reward value, new state) into the experience replay pool; S5.6 periodically samples small batches of experience and uses DQN to update network weights; the target network is updated every certain number of steps.
[0032] The DQN reinforcement learning model training process incorporates mechanisms such as experience replay and delayed target network updates to improve convergence stability and prevent the policy from getting trapped in local optima. The corresponding network flowchart is shown below. State S is defined as the environmental space composed of controlled and disturbance variables of the data center cooling system, including air conditioning supply air temperature setpoint, air conditioning return air temperature setpoint, air conditioning on / off status, server room temperature, fan speed, IT equipment load, and outdoor temperature. A is defined as the action space composed of control variables, including the air conditioning supply and return air temperature setpoints and the air conditioning on / off status. R is defined as the reward signal, i.e., the reward value obtained by reinforcement learning when the air conditioning in the data center is in state S and performs action A. The action network and evaluation network are implemented using a multilayer perceptron structure. The action network is responsible for selecting the next action A to be performed based on the current state S. The evaluation network provides a value estimate q(s,a) for the current state S and action A, used to optimize the controller's action selection, select the most suitable action A output, and execute and update it in the data center environment.
[0033] III. Input Feature Correlation Analysis
[0034] To improve the model's generalization ability and training efficiency, this paper performs Pearson correlation analysis on the input features during the model construction phase. Variables strongly correlated with future temperature predictions are selected, including key supply air temperatures, historical averages of specific temperature points, and temperature change rates, while low-correlation features are removed to avoid redundancy. Based on this result, the simplified input features significantly reduce model complexity, improve the convergence speed and generalization performance of the XGBoost model, and also help the reinforcement learning strategy make reasonable decisions with clearer state information.
[0035] The inventors discovered through experience that the changing trends of some input characteristic quantities (state physical quantities) exhibit a high degree of consistency, such as... Figure 2 As shown. Directly inputting temperature data as independent variables into the model can easily lead to information duplication and even overfitting. Therefore, we introduced an "expert identification" mechanism by integrating feature engineering and structural prior knowledge, and combined it with feature dimensionality reduction methods to construct more representative temperature input features, thereby improving the model's prediction accuracy and training efficiency. During this process, because the computer room temperature control is affected by outdoor temperature, day-night temperature, or seasonal changes, the air conditioning system needs to make strategic adjustments to maintain stability. We transformed knowledge from industry experts and engineers into structured input features, i.e., "expert identification," which guides the model to focus on key influencing factors of the temperature control system. This step is mainly used to screen temperature points sensitive to outdoor temperature changes, temperature zones (multiple temperature points) with control correlations with specific air conditioning units, and high-risk areas under temperature control imbalance. After completing the aggregation and dimensionality reduction of temperature points, we further analyzed the correlation between regional temperature features and each air conditioning unit. The specific steps are as follows: First, the environmental data is standardized preprocessed to obtain standardized feature vectors. The standardized preprocessing includes missing value processing, outlier removal, and normalization. Then, based on prior knowledge of feature engineering and engineering structure, indoor temperature points closely related to outdoor temperature, air conditioning parameters and dynamic indicators are recommended through expert identification; (this step is used to screen temperature points sensitive to changes in outdoor temperature, temperature zones (multiple temperature points) with control association with specific air conditioning units, and high-risk areas with temperature control imbalance).
[0036] S1.3, Based on Principal Component Analysis (PCA), eigenvalue vectors are obtained for each temperature point according to their covariance, and then sorted according to the eigenvalues. The following formula is used to ensure that the cumulative covariance is greater than a preset threshold. Identify the minimum required features in time. , In the formula, x j Let j represent the j-th feature, k represent the total number of features, and f represent the cumulative covariance ratio function. A larger function value indicates a better fit. This indicates a preset threshold, ranging from 85% to 95%.
[0037] The input features are reduced in dimensionality by using an "expert identification" mechanism and correlation analysis, thereby accelerating convergence: the number of training rounds is reduced to 200.
[0038] Example
[0039] Taking the third-floor computer room as an example, the effectiveness of the algorithm strategy of this invention is verified by comparing air conditioning energy consumption, energy saving rate, exceedance rate, and PUE value through different control strategy algorithms. The algorithm strategy of this invention and traditional PID control were compared using data from the computer room's operation from April to July 2025. The results are shown in Tables 1 and 2. Table 1 compares the daily power consumption of the computer room extracted by different control methods, and Table 2 compares the power consumption of the 7-day test set of different control methods.
[0040] Table 1
[0041] Table 2
[0042] As shown in Table 1, the energy consumption of the same strategy varies significantly under different outdoor temperatures, with an overall positive correlation to outdoor temperature. Compared to PID and other strategies, the LSTM-DQN model requires a large number of training samples, has high inference latency, relies on high-precision thermodynamic parameters in its physical model, and exhibits poor generalization. Meanwhile, the conventional Pure DQN model typically relies on real-time data for data center control, lacking the forward-looking advantage of a prediction module. After screening multiple models, this invention's model achieved the highest energy savings of 5.52%, 6.36%, and 9.34% under different outdoor temperatures, combining efficiency and accuracy. Its measured training speed is three times faster than LSTM, making it more suitable for the rapid iteration requirements of data center models. Table 2 further shows that the control strategy of this invention has the lowest energy consumption and PUE value, and the environmental temperature alarm exceedance rate is as low as 0.21%.
[0043] Figure 3 The graph shows the hourly power consumption of the air conditioner on April 6th and 7th, 2025, using PID control and the control mode of this invention, respectively. As can be seen from the graph, compared to traditional PID control, the hourly power consumption of the air conditioner under the control strategy of this invention is generally lower than that of the conventional control mode (the peak data in the graph is due to the triggering of the safety net strategy).
[0044] Figure 4 This is a summary comparison of the power consumption of the computer room air conditioning in April 2025 using PID and other strategies, compared with the control mode of this invention. From... Figure 4 It can be seen that: under the traditional mode (PID and other strategies), the running time accounts for 52%, with a total power consumption of 64960.4 kWh; under the present invention, the running time accounts for 48%, with a total power consumption of 59189.0 kWh, and the average energy saving rate in April is 6.00%. Based on the data center's daily raw power consumption of approximately 1200 kWh, and calculating with an energy saving rate of 8%, the estimated daily power saving is 96 kWh. Therefore, the annual power saving for the third-floor data center alone would be 96 * 365 = 35040 kWh.
[0045] This invention is the first to embed XGBoost predictions into the RL state space, enabling pre-adjustment based on future temperature risks. It also integrates domain knowledge to construct an air conditioning-temperature zone correlation matrix and combines PCA dimensionality reduction to compress feature dimensions by up to 70%. Deployed in a data center with over 2000 racks, it has achieved measured energy savings exceeding 12%. Experiments demonstrate that this invention significantly improves energy efficiency while ensuring data center security, providing a new approach to green data center development.
[0046] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A data center air conditioning energy-saving control method based on joint prediction and reinforcement learning, characterized in that... Includes the following steps: S1, Collect historical data: Collect environmental data of the data center computer room during historical periods, including: time information, location temperature, outdoor temperature, air conditioning parameters and dynamic indicators; S2, Prediction Model Training: A data center temperature prediction model is built based on the XGBoost algorithm, and the historical data of the data center collected in S1 is used as training samples to complete the model training and optimization, and obtain the trained data center temperature prediction model. S3 acquires the current environmental data of the data center and, based on the model trained in S2, uses the current environmental data as input features to output the temperature prediction result through the optimized temperature prediction model. S4. Embed the temperature prediction results obtained in step S3 into the environmental data of the data center to obtain the data center state S. By incorporating the prediction information into the state space, DQN can make forward-looking decisions. S5. Construct a DQN reinforcement learning model and a reward function R. Take the data center state S and the control quantity A of the air conditioning unit action execution as input data, and output the updated control quantity A' of the air conditioning unit action execution. The updated control quantity A' of the air conditioning unit action execution is applied to the data center to obtain the updated data center PUE value and reward value R. S6. Repeat steps S4 and S5 to perform multiple rounds of iterative training and optimization on the DQN reinforcement learning model. When the data center PUE value and the reward function feedback value R tend to stabilize, the trained DQN reinforcement learning model is obtained. S7. Based on the DQN reinforcement learning model trained in step S6, the data center state S, which incorporates the temperature prediction results from the XGBoost algorithm, is used as input data to complete the output of the data center air conditioning energy-saving control quantity A.
2. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 1, characterized in that: In step S1, the air conditioning parameters include supply air temperature, return air temperature, fan speed, fan frequency, and compressor frequency; the dynamic indicators include temperature change rate and IT load power change; in step S2, the temperature prediction result is the maximum temperature of the data center in the next 5 minutes. .
3. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 1, characterized in that: In step S2, the XGBoost algorithm uses a serial CART regression tree to progressively optimize the loss function to predict the ambient temperature of the data center. The constructed loss function L is as follows: ， In the formula, This represents the true temperature value of the i-th sample. This represents the predicted temperature value for the i-th sample. The tree complexity regularization parameter is used to control splitting, and its value is 0.
1. This represents the total number of CART regression trees. This represents the L2 regularization coefficient, with a value of 0.
1. The L2 norm represents the weight of the leaf node.
4. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 1, characterized in that: The reward function R described in step S4 is designed as follows: ， In the formula, T represents the maximum temperature of the data center in the next 5 minutes. threshold The temperature alarm threshold is set to 25℃, E represents the data center energy consumption, α represents the weighting coefficient for the balanced temperature control progress, and β represents the weighting coefficient for energy consumption optimization.
5. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 1, characterized in that: In step S1, the sample data undergoes dimensionality reduction processing as follows: S1.1, Standardize the environmental data to obtain a standardized feature vector. The standardization preprocessing includes missing value processing, outlier removal, and normalization processing. S1.2, based on feature engineering and prior knowledge of engineering structure, optimizes and recommends indoor temperature points that are closely related to outdoor temperature, air conditioning parameters and dynamic indicators through expert identification; S1.3, Based on Principal Component Analysis (PCA), eigenvalue vectors are obtained for each temperature point according to their covariance, and then sorted according to the eigenvalues. The following formula is used to ensure that the cumulative covariance is greater than a preset threshold. Identify the minimum required features in time. ， In the formula, x j Let j represent the j-th feature, k represent the total number of features, and f represent the cumulative covariance ratio function. A larger function value indicates a better fit. This indicates a preset threshold, ranging from 85% to 95%.
6. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 4, characterized in that: The iterative training and optimization process of the DQN reinforcement learning model described in step S5 is as follows: S5.1, Initialize the DQN reinforcement learning model policy and target network; S5.2, At the current time step, based on the current state, use the ε algorithm to determine the control quantity of the action to be executed in the action network; S5.3 After executing action A on the computer room air conditioner according to the control quantity, update the air conditioning unit parameter settings and predict the temperature data for the next five minutes through the XGBoost temperature prediction model, and obtain the maximum temperature value T in the predicted temperature. S5.4 Construct a new data center state S, calculate the reward value of the action based on the reward function R, and feed the reward value of the action back to the DQN reinforcement learning model to update the model policy. S5.5, Repeat steps S5.2-S5.4 until the reward value of the action tends to stabilize, and complete the iterative training and optimization of the model.
7. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 6, characterized in that: Construct an experience replay pool and store the data center state S, the action A, the reward value R, and the new data center state S' obtained in step S5.5 for each time step into the experience replay pool. Every certain number of time steps, samples are taken from the experience replay pool, and the model weights are updated using DQN to complete the periodic update of the target network.
8. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 6, characterized in that: In step S6, the iterative training and optimization of the DQN reinforcement learning model ends when the following condition is met. ， In the formula, This represents the reward value in the j-th iteration. This represents the reward value in the (j+n)th iteration. It is the constraint coefficient, with a value range of 0.01-0.
1.
9. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 6, characterized in that: By performing Pearson correlation analysis on the input features of the XGBoost temperature prediction model, variables strongly correlated with future temperature prediction were selected. These strongly correlated variables include the air conditioning supply temperature setpoint, the location temperature, the average location temperature over a 20-minute period, and the temperature change rate.
10. The data center air conditioning energy-saving control method based on joint prediction and reinforcement learning according to claim 6, characterized in that: When iteratively training the DQN reinforcement learning model The action network uses a multilayer perceptron (MLP) with a structure consisting of a 64-node input layer and a 32-node hidden layer, both of which use the ReLU activation function. The output layer dimension is the action space size |A|. The target network is a copy of the action network, and is trained stably through a delayed update mechanism with a delayed update coefficient τ=0.01.