Machine room air conditioner group control method, device, equipment, storage medium and program product

Abstract

The invention discloses a group control method, device and equipment for air conditioners in a machine room, a storage medium and a program product, and relates to the technical field of environment regulation, and the group control method for the air conditioners in the machine room comprises the steps that current operation data of different air conditioners in the machine room and current associated environment data of associated areas corresponding to all the air conditioners are obtained; for the target air conditioner, the current local observation vector of the target air conditioner is determined according to the current operation data and the current associated environment data; historical actions of the target air conditioner at the last moment are obtained, and the current local action value of the target air conditioner is determined according to the historical actions and the current local observation vector; and determining a global joint value of the machine room according to the local action value of each air conditioner, and determining an air conditioner group control scheme of the machine room according to the global joint value. By maximizing the global joint value, all the air conditioner intelligent agents can be guided to adjust actions in the global optimal direction, air conditioner group control of the machine room is achieved, and the problems of isolated island operation and lack of cooperation are avoided.

Description

Technical Field

[0001] This application relates to the field of environmental control technology, and in particular to methods, devices, equipment, storage media and program products for group control of computer room air conditioning. Background Technology

[0002] With the explosive growth of cloud computing, big data, and artificial intelligence, the energy consumption problem of data centers is becoming increasingly serious. Heating, ventilation, and air conditioning (HVAC) systems typically account for 30%-40% of total energy consumption, making them the largest energy-consuming unit after information equipment. How to reduce cooling energy consumption (lower PUE value) through efficient control strategies while ensuring the safe operating temperature of IT equipment is a pressing problem that needs to be solved in the current technological field.

[0003] Early and currently, most small and medium-sized data centers still use single-unit control based on Proportional-Integral-Derivative (PID) logic or simple group control logic. Each computer room air conditioner (CRAC) independently adjusts its temperature setting based solely on the temperature sensor readings at its own return air vent. This approach is essentially "island operation," lacking coordination between devices. Summary of the Invention

[0004] The main purpose of this application is to provide a method, device, equipment, storage medium and program product for group control of computer room air conditioners, which aims to solve the technical problem of existing computer room air conditioners that independently adjust temperature settings and lack coordination between devices.

[0005] To achieve the above objectives, this application proposes a method for group control of computer room air conditioning, the method comprising: Obtain the current operating data of different air conditioners in the computer room and the current associated environmental data of the corresponding associated areas of each air conditioner; For a target air conditioner, the current local observation vector of the target air conditioner is determined based on the current operating data and the current associated environmental data; wherein, the target air conditioner is any air conditioner in the computer room; Obtain the historical actions of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector; The global combined value of the computer room is determined based on the local action value of each air conditioner, and the air conditioning group control scheme of the computer room is determined based on the global combined value.

[0006] In one embodiment, the step of determining the global joint value of the computer room based on the local action value of each of the air conditioners includes: Construct a joint input vector based on the local action value of each air conditioner; The joint input vector is corrected by a multi-head self-attention module to obtain a corrected output vector; wherein, the corrected output vector includes the corrected value of each air conditioner, and the corrected value is the value after correcting the local action value; The corrected values ​​of each air conditioner are nonlinearly aggregated based on the topology-aware hybrid network module to obtain the global joint value of the computer room.

[0007] In one embodiment, the multi-head self-attention module includes: three independent linear transformation units, a multi-head self-attention unit, a residual connection unit, and an output unit; The step of correcting the joint input vector using a multi-head self-attention module to obtain a corrected output vector includes: The joint input vector is input into three independent linear transformation units for linear projection to obtain the feature representation of the joint input vector. The feature representation is input into the multi-head self-attention unit to perform multi-head attention calculation, and attention weights are obtained. The attention weights are residually connected using the residual joint unit to obtain the fused features; The fused features are layer-normalized using the output unit to obtain a corrected output vector.

[0008] In one embodiment, the topology-aware hybrid network module includes: a graph neural network state encoder, a hybrid network, and a supernetwork corresponding to each layer of the hybrid network; The step of performing nonlinear aggregation of the corrected values ​​of each air conditioner based on the topology-aware hybrid network module to obtain the global joint value of the computer room includes: Obtain the graph structure data corresponding to the computer room, and extract features from the graph structure data through the graph neural network state encoder to obtain a global topology embedding vector; The global topology embedding vector is input into the hypernetwork to determine the hypernetwork weights and hypernetwork biases for each layer of the hypernetwork. The global joint value of the data center is obtained by weighting the modified value using the hybrid network based on the supernetwork weights and the supernetwork biases.

[0009] In one embodiment, before the step of obtaining the historical actions of the target air conditioner at the previous moment and determining the current local action value of the target air conditioner based on the historical actions and the current local observation vector, the method further includes: Obtain an offline training dataset and perform reinforcement learning based on the offline training dataset to obtain a pre-trained model; Based on the pre-trained model, simulation training is performed to obtain the simulation training model; The simulation training model is fine-tuned for a real environment to obtain a real environment model; the real environment model includes at least a multi-head self-attention module and a topology-aware hybrid network module.

[0010] In one embodiment, the state space of the reinforcement learning is determined based on the operating indicators of the computer room and graph structure data, the action space of each air conditioner is the set of control instructions corresponding to the air conditioner, and the reward function is a multi-objective composite reward function.

[0011] Furthermore, to achieve the above objectives, this application also proposes a data center air conditioning group control device, which includes: The data acquisition module is used to acquire the current operating data of different air conditioners in the computer room and the current associated environmental data of the associated area corresponding to each air conditioner; A vector construction module is used to determine the current local observation vector of a target air conditioner based on the current operating data and the current associated environmental data; wherein the target air conditioner is any air conditioner in the computer room; The local value management module is used to obtain the historical actions of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector. The global management module determines the global combined value of the computer room based on the local action value of each air conditioner, and determines the air conditioning group control scheme of the computer room based on the global combined value.

[0012] In addition, to achieve the above objectives, this application also proposes a data center air conditioning group control device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the data center air conditioning group control method described above.

[0013] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the computer room air conditioning group control method described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the data center air conditioning group control method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: This application achieves regional division using physical topology by collecting the operating parameters of each air conditioner and data from its associated areas. This allows each air conditioner agent to focus only on its local environment, avoiding the spatial information loss caused by the flattening and splicing of hundreds or thousands of sensors in traditional methods. This lays the foundation for subsequent graph neural network modeling and decentralized execution. By determining the current local observation vector based on the air conditioner's current operating data and the current associated environmental data, each air conditioner agent can accurately perceive the heat load of its associated area and its own operating status, thus coupling decision-making with physical reality. By maximizing the global joint value, all air conditioner agents can be guided to adjust their actions in a globally optimal direction, realizing group control of air conditioners in the data center and avoiding the problems of isolated operation and lack of coordination. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the data center air conditioning group control method of this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the data center air conditioning group control method of this application; Figure 3 This is a G-QMIX network structure diagram in one implementation of the data center air conditioning group control method of this application; Figure 4 This is a flowchart illustrating Embodiment 3 of the data center air conditioning group control method of this application; Figure 5 This is a flowchart illustrating the three-stage transfer training process in one implementation of the data center air conditioning group control method of this application. Figure 6 This is a schematic diagram of the basic reinforcement learning structure in one implementation of the data center air conditioning group control method of this application; Figure 7This is a schematic diagram of the modular structure of the data center air conditioning group control device according to an embodiment of this application; Figure 8 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the group control method for computer room air conditioning in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of this application embodiment is as follows: obtain the current operating data of different air conditioners in the computer room and the current associated environmental data of the corresponding associated areas of each air conditioner; for the target air conditioner, determine the current local observation vector of the target air conditioner based on the current operating data and the current associated environmental data; wherein, the target air conditioner is any air conditioner in the computer room; obtain the historical action of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical action and the current local observation vector; determine the global joint value of the computer room based on the local action value of each air conditioner, and determine the air conditioner group control scheme of the computer room based on the global joint value.

[0023] In some embodiments of this application, a Computational Fluid Dynamics (CFD) model of the data center can be established to predict future temperature field changes, and the air conditioning can be controlled by solving optimization equations. This scheme addresses the coordination problem between devices by introducing Model Predictive Control (MPC), but it heavily relies on the accuracy of the physical model. Establishing a high-precision real-time CFD model is extremely difficult, and the model requires frequent calibration as the rack utilization rate changes. Furthermore, CFD computation is enormous, making it difficult to meet the requirements of real-time control.

[0024] In some embodiments of this application, since reinforcement learning (RL) does not require prior knowledge of the physical equations of the environment, but learns the optimal policy through trial and error with the environment, air conditioning control can be achieved through reinforcement learning. However, as the size of the computer room increases (e.g., 20 air conditioners, each with multiple actions), the joint action space explodes exponentially, causing the algorithm to fail to converge.

[0025] In some embodiments of this application, a group control method of Graph-AttentionalQMIX (G-QMIX) and its supporting three-stage training strategy of "offline pre-training - simulation reinforcement training - online fine-tuning" are also proposed. In the embodiments of this application, the core idea of ​​G-QMIX is to model the state of the computer room as a graph structure, use a graph neural network (GNN) to extract the topological features of the graph structure, and introduce a self-attention mechanism to break the monotonicity of value decomposition, thereby realizing improved air conditioning control that integrates physical topological priors and has the ability to express complex strategies.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or virtual device capable of performing the above functions. The following description uses a data center air conditioning group control device (hereinafter referred to as the group control device) as an example to illustrate this embodiment and the subsequent embodiments.

[0027] Based on this, the embodiments of this application provide a method for group control of computer room air conditioning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the data center air conditioning group control method of this application.

[0028] In this embodiment, the data center air conditioning group control method includes steps S10~S40: Step S10: Obtain the current operating data of different air conditioners in the computer room and the current associated environmental data of the associated area corresponding to each air conditioner.

[0029] It should be noted that a power environment monitoring and management system can be deployed in the computer room of this application embodiment. This system can collect data from the computer room in real time at a certain sampling frequency. When sampling, the sampling frequency can be set according to the needs of actual application, such as sampling once per minute, sampling once every half minute, etc., and this application embodiment does not impose any restrictions on this.

[0030] In this embodiment of the application, the power environment monitoring and management system can be used to collect air conditioning operation data, cabinet environment data, global environment data, etc., and this embodiment of the application does not limit it.

[0031] It should be noted that air conditioner operation data may include the on / off status of each air conditioner (S). on / off ), set temperature (T) set ), return air temperature (T) ret ), fan speed (S) fan ), air conditioning operating power (P) run The embodiments of this application do not limit this.

[0032] It should be explained that each air conditioner in the data center of this application can be associated with several sensors within its air supply area. This sensor data constitutes the rack environment data associated with the air conditioner. These sensors can be temperature and humidity sensors or other sensors used to collect temperature values ​​from the hot and cold aisles. and humidity value ,in Indicates the number of sensors.

[0033] It is understandable that the aforementioned global environmental data may specifically include outdoor ambient temperature (T). amb Total IT load power of the computer room (P) total_load ), average temperature of the computer room (T) avg ), average humidity of the computer room (H avg The embodiments of this application do not limit this.

[0034] It should be understood that the aforementioned current operating data of different air conditioners refers to the air conditioner operating data of different air conditioners in the computer room collected at the current moment. The aforementioned current associated environmental data refers to the server rack environmental data associated with the air conditioners in the computer room collected at the current moment.

[0035] It should be noted that, based on the above data, the collected discrete sensor time-series data can be transformed into graph-structured data that the algorithm can understand. The graph structure in this application embodiment may include control nodes and environment nodes.

[0036] In some embodiments of this application, the solution is illustrated by taking four air conditioners installed in the computer room as an example. In practical applications, the number of air conditioners installed in the computer room may be more or less than four, such as three, five, or ten. This application does not limit the number of air conditioners installed in the computer room.

[0037] It should be noted that the above control node can be represented as These correspond to four air conditioners. The feature vector of each control node... After normalization, it can be expressed as The above environment nodes can be represented as: This corresponds to multiple thermally critical points (such as hot and cold aisle sensors) and equipment within the computer room. The feature vector of each environmental node... After normalization, it can be expressed as ,in This indicates the load power of the environmental node.

[0038] It should be explained that in the graph structure, control nodes and environment nodes can be connected by directed edges. In this embodiment, directed edges can be established based on the thermal influence domain of physical space. For example, a threshold distance to the forward air vent of an air conditioner can be set. Meters. If the server rack With air conditioning Straight-line distance between forward air outlets Then directed edges can be established. For adjacent air conditioners and air conditioning If the distance between the two (in, The value can be chosen based on the actual application, such as 0.5. Then, a two-way edge representing the presence of airflow interference can be established. The weights corresponding to each edge can be inversely proportional to the distance, such as... ,in To represent a constant, used to avoid a zero denominator, it is usually a very small positive number, such as... .

[0039] It is understandable that the associated area corresponding to the air conditioner can be identified as the equipment or sensor corresponding to the environmental node or control node that has an edge connection with the air conditioner.

[0040] In practice, the group control device acquires the current operating data of different air conditioners in the computer room and the current associated environmental data of the associated area of ​​each air conditioner, so that the intelligent agent corresponding to the air conditioner can know the current operating status and decide the next action based on it.

[0041] Step S20: For the target air conditioner, determine the current local observation vector of the target air conditioner based on the current operating data and the current associated environmental data; The target air conditioner can be any air conditioner in the computer room.

[0042] It should be noted that any air conditioner in the computer room can correspond to an air conditioner agent, which can be equipped with a deep recurrent Q-network. A deep recurrent Q-network (DRQN) is a model that combines deep learning, recurrent neural networks, and the Q-learning algorithm from reinforcement learning. Compared to traditional deep Q-networks, by introducing recurrent neural network layers, it enhances the network's memory capacity, inferring information that cannot be directly seen at present based on past observations.

[0043] In this embodiment, the current local observation vector is an observation vector determined based on the air conditioner's current operating data and the current associated environmental data of the associated region. The local action value of the air conditioner agent in the current state can be determined based on this current local observation vector and the air conditioner's historical actions at the previous moment.

[0044] That is, in step S30, the historical actions of the target air conditioner at the previous moment are obtained, and the current local action value of the target air conditioner is determined based on the historical actions and the current local observation vector.

[0045] It should be noted that the aforementioned historical actions refer to the control commands actually executed by the air conditioner at the previous control moment. In this embodiment, the air conditioning optimization problem is modeled as a decentralized partially observable Markov decision process. In partially observable environments, since a single air conditioning agent cannot obtain the complete state of the environment, it needs to rely on historical information to infer the current state. The historical actions at the previous moment are an important part of the historical information, which can help the air conditioning agent understand the control commands at the previous moment and the corresponding responses, such as lowering the set temperature at the previous moment, but the ambient temperature did not decrease.

[0046] It should be understood that the aforementioned local action value can be an estimate of the future cumulative reward for each possible action based on the current hidden state. The deep recurrent Q-network in this embodiment may include an input layer, a memory layer, and an output layer. Both the input and output layers can be multilayer perceptrons (MLPs), and the memory layer can be a gated recurrent unit (GRU).

[0047] In this embodiment, the local observation vector is... Historical actions of the previous moment The input is fed into a gated loop unit, which processes the timing information and updates the hidden layer state. The hidden state is obtained by continuously accumulating historical observations and actions, implicitly encoding the entire history. By processing the updated hidden state through the output layer, the value estimate of each candidate action can be output, thus obtaining the local action value. .

[0048] It should be noted that the action space of the air conditioner intelligent agent in this application embodiment can be set according to the corresponding control commands of the air conditioner, and this application embodiment does not limit this. For example, the size of the action space... This includes one shutdown action and 13 power-on temperature setting levels.

[0049] This application's embodiments determine the current local observation vector of the target air conditioner based on current operating data and current associated environmental data; obtain the target air conditioner's historical actions from the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector. Compared to the lag in current error adjustment provided by PID control, this application, by introducing the historical actions from the previous moment, enables the air conditioner's intelligent agent to perceive the lag effect of actions. Linking the local action value with the historical actions from the previous moment enables learning of frequent switching behaviors, avoiding energy waste and shortened lifespan caused by frequent start-stop of the air conditioner.

[0050] Step S40: Determine the global combined value of the computer room based on the local action value of each air conditioner, and determine the air conditioning group control scheme of the computer room based on the global combined value.

[0051] Understandable is the global joint value. In multi-agent reinforcement learning, the global joint value is used to evaluate the overall expected cumulative reward of all air conditioning agents under certain joint actions and joint historical trajectories. It reflects the total future benefit that the data center can obtain after taking this set of control commands in the current state. In this embodiment, the global joint value can be obtained by linearly summing the local action values ​​of each air conditioning agent, or by using monotonic hybrid networks, weight repair, etc. This embodiment does not impose any limitations on this method. By maximizing the global joint value, all air conditioning agents can be guided to adjust their actions towards the global optimum, achieving group control of the data center's air conditioning systems and avoiding the problems of isolated operation and lack of coordination.

[0052] This application embodiment acquires the current operating data of different air conditioners in the computer room and the current associated environmental data of the corresponding areas of each air conditioner; for a target air conditioner, the current local observation vector of the target air conditioner is determined based on the current operating data and the current associated environmental data; the historical actions of the target air conditioner at the previous moment are acquired, and the current local action value of the target air conditioner is determined based on the historical actions and the current local observation vector; the global joint value of the computer room is determined based on the local action value of each air conditioner, and the air conditioner group control scheme of the computer room is determined based on the global joint value. By collecting the operating parameters of each air conditioner and the data of its associated areas, the region division using physical topology is realized, so that each air conditioner agent only needs to focus on the local environment, avoiding the spatial information loss caused by the flat splicing of hundreds or thousands of sensors in traditional methods, laying the foundation for subsequent graph neural network modeling and decentralized execution. By determining the current local observation vector based on the current operating data of the air conditioner and the current associated environmental data, each air conditioner agent can accurately perceive the heat load of the associated area and its own operating status, so that the decision-making basis is coupled with the physical reality. By maximizing the global joint value, all air conditioning agents can be guided to adjust their actions in the direction of global optimization, realizing group control of air conditioning in the computer room and avoiding the problems of isolated operation and lack of coordination.

[0053] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2 This is a flowchart illustrating Embodiment 2 of the data center air conditioning group control method of this application.

[0054] like Figure 2 As shown in the embodiment of this application, the step of determining the global joint value of the computer room based on the local action value of each air conditioner includes: Step S41: Construct a joint input vector based on the local action values ​​of each air conditioner; Step S42: The joint input vector is corrected by a multi-head self-attention module to obtain a corrected output vector; The corrected output vector includes the corrected value of each air conditioner, and the corrected value is the value after correcting the local action value; Step S43: Based on the topology-aware hybrid network module, the corrected values ​​of each air conditioner are nonlinearly aggregated to obtain the global joint value of the computer room.

[0055] It should be noted that in this implementation, the G-QMIX network may include a three-layer hierarchical structure: a bottom-layer local decision network, a middle-layer non-monotonic interaction layer, and a top-layer topology-aware hybrid network. Specifically, it can be as follows: Figure 3As shown, Figure 3 This is a G-QMIX network structure diagram in one implementation of the data center air conditioning group control method of this application.

[0056] Understandably, the underlying local decision-making network can be composed of air conditioning agents (Agent 1, Agent 2, ..., Agent n) as described above. Each air conditioning agent can be equipped with a deep recurrent Q-network, and the local action value of the air conditioning agent in the current state is output based on the deep recurrent Q-network. .

[0057] For example, the current state of the air conditioning intelligent agent 1 at the current time t. The corresponding local action value is ,in This indicates the actions of the air conditioning intelligent agent, Agent 1. This represents the historical information of the air conditioning agent 1, such as the hidden layer state; and the current state of the air conditioning agent n at the current time t. The corresponding local action value is .

[0058] It should be noted that the non-monotonic interaction layer in the middle layer is located between the local decision network and the topology-aware hybrid network. By setting up a multi-head self-attention (MHSA) module, the team context information of each air conditioner in the computer room is corrected, thereby determining the corrected value of each air conditioner. This breaks the strict monotonic correspondence between local value and global value, allowing the model to express complex collaborative logic.

[0059] It is understandable that the aforementioned joint input vector is a vector composed of the local action values ​​of each air conditioner. For example, the joint input vector composed of the local action values ​​of air conditioner agents Agent 1, ..., Agent n is: , This represents the transpose of a matrix.

[0060] It should be noted that by introducing a multi-head self-attention mechanism between the local action value and the topology sensing network through a multi-head self-attention module, the local action values ​​of all air conditioning agents can be mapped to high-dimensional vectors, thus determining the corrected local action values. Specifically, this can be illustrated as follows: ; ; in, This represents the corrected local action value of the air conditioning agent Agent i (i=1, 2, ..., n). This represents the value of the local actions of the air conditioning intelligent agent (Agent i) before the correction. Indicates attention weights. express The high-dimensional vector obtained by mapping This represents the local action value of the air conditioning intelligent agent j. The high-dimensional mapping vector j traverses all intelligent agents. This represents the query weight matrix. Represents the key weight matrix. Value weight matrix. The scaling factor can be used to select the dimension of the key. This represents a linear transformation process (such as a linear transformation in the output layer) to input fused features. The fused features, obtained by attention-weighted fusion, are then added element-wise to the original features before being processed. Perform layer normalization. This is the normalization function. This represents the query matrix, i.e., the feature representation of the linear layer Q; The key matrix is ​​the feature representation of the linear layer K. It is the value matrix, that is, the feature representation of the linear layer V.

[0061] Specifically, the multi-head self-attention module in this application embodiment may include: three independent linear transformation units (Linear), a self-attention unit (Self-Attention unit, specifically an h-head multi-head self-attention unit), a residual connection unit (Concat), and an output unit (Linear). The step of correcting the joint input vector using the multi-head self-attention module to obtain a corrected output vector includes: inputting the joint input vector into the three independent linear transformation units for linear projection to obtain a feature representation of the joint input vector; inputting the feature representation into the multi-head self-attention unit for multi-head attention calculation to obtain attention weights; performing residual connection on the attention weights through the residual connection unit to obtain fused features; and performing layer normalization on the fused features through the output unit to obtain the corrected output vector.

[0062] It is understandable that the three independent linear transformation units mentioned above are linear layer V, linear layer K, and linear layer Q. The final corrected output vector can be expressed as... ,in That is, the local value vector of the air conditioning agent Agent 1 after correction, and the others are similar.

[0063] It should be noted that by weighting through attention weights, the corrected local action value incorporates the global collaborative context, allowing the corrected local action value to remain high even when the local action value decreases.

[0064] In some embodiments of this application, the topology-aware mixing network of this application enables GQMIX to no longer rely on global state, but instead utilize the spatial topology structure of the environment (i.e., graph structure data). The parameters of the hybrid network are generated to enable the aggregation process of global value to perceive the physical distance and connection relationship between air conditioning agents. Specifically, the topology-aware hybrid network module includes: a graph neural network state encoder, a hybrid network, and a supernetwork corresponding to each layer of the hybrid network; the step of nonlinearly aggregating the corrected values ​​of each air conditioner based on the topology-aware hybrid network module to obtain the global joint value of the computer room includes: acquiring graph structure data corresponding to the computer room, and extracting features from the graph structure data through the graph neural network state encoder to obtain a global topology embedding vector; inputting the global topology embedding vector into the supernetwork to determine the supernetwork weight and supernetwork bias corresponding to each layer of the supernetwork; and using the hybrid network to perform a mixed weighting of the corrected value based on the supernetwork weight and supernetwork bias to obtain the global joint value of the computer room.

[0065] It should be noted that the graph neural network state encoder in this embodiment uses a two-layer graph neural network (which may be GAT / GCN) to process graph structure data (i.e., input projection) to obtain a global topological embedding vector. (i.e., output projection), for example: ; in, This indicates a global average pooling operation. This represents the graph structure data at the current time t. This represents the node feature matrix at the current time t. This represents graph neural network processing.

[0066] It should be noted that the parameters of the hybrid network (i.e., the hypernetwork weights and biases) are generated by inputting the global topology embedding vector into the hypernetwork. For each layer of the hybrid network, there can be a corresponding hypernetwork, which is typically a simple feedforward network (e.g., a single-layer linear network with an L+ absolute value activation function |·|) to ensure that the generated weights are non-negative, such as: ; ; in, For the first The hypernetwork weights corresponding to the layers of the hypernetwork. For the first The supernetwork bias corresponding to each layer of the supernetwork. The weights / biases of each supernetwork layer are fused layer by layer.

[0067] In this embodiment of the application, a two-layer hypernetwork is used as an example to illustrate the formula for the output of the global joint value of the application, but this does not constitute a limitation on the number of hypernetwork layers actually used in the application.

[0068] ; in, For the sake of overall joint value, , These represent the supernetwork weights corresponding to the first and second layers of the supernetwork, respectively. and These represent the hypernet biases corresponding to the first and second layers of the hypernet, respectively. This is the activation function.

[0069] This embodiment constructs a joint input vector based on the local action value of each air conditioner; the joint input vector is then corrected using a multi-head self-attention module to obtain a corrected output vector; the corrected output vector includes the corrected value of each air conditioner, which is the value after correcting the local action value; the corrected values ​​of each air conditioner are nonlinearly aggregated using a topology-aware hybrid network module to obtain the global joint value of the computer room. Because the local action value is corrected using a multi-head self-attention module, the monotonicity constraint of traditional QMIX is broken, enabling the model to learn complex cooperation strategies. Each air conditioner agent can pay attention to the state of other related agents while estimating its own local action value, implicitly realizing information sharing among air conditioner agents and avoiding the inefficiency of requiring extensive trial and error to discover cooperative relationships. The nonlinear aggregation of the corrected values ​​of each air conditioner using a topology-aware hybrid network allows the value aggregation process to perceive the spatial distance and airflow coupling relationship between air conditioners, enabling the model to physically understand which air conditioners should cooperate, thus improving the rationality of the decision.

[0070] Based on the first and / or second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to the first and / or second embodiments described above can be referred to the above description and will not be repeated hereafter. Based on this, please refer to... Figure 4 , Figure 4 This is a flowchart illustrating Embodiment 3 of the data center air conditioning group control method of this application.

[0071] like Figure 4 As shown in the embodiment of this application, before the step of obtaining the historical actions of the target air conditioner at the previous moment and determining the current local action value of the target air conditioner based on the historical actions and the current local observation vector, the method further includes: Step S100: Obtain the offline training dataset and perform reinforcement learning based on the offline training dataset to obtain the pre-trained model; Step S200: Perform simulation training based on the pre-trained model to obtain a simulation training model; Step S300: Fine-tune the simulation training model to obtain a real environment model. The real-world environment model includes at least: a multi-head self-attention module and a topology-aware hybrid network module.

[0072] It should be noted that, to address the cold start risk and environmental adaptability issues in engineering implementation, this application's embodiments design a three-stage transfer training strategy of "offline-simulation-online". Specifically, it can be described as follows... Figure 5 As shown, Figure 5 This is a flowchart illustrating the three-stage migration training process in one implementation of the data center air conditioning group control method of this application.

[0073] Reference Figure 5 In this embodiment, the offline training phase is the pre-training phase based on historical data (Phase 1: Historical Data Training). Specifically, this embodiment can acquire historical operational data from the computer room (including state-action trajectories, etc.) as the offline training dataset, and use Behavior Cloning (BC) and offline reinforcement learning algorithms to pre-train the G-QMIX network to obtain a pre-trained model. The offline training phase aims to enable the air conditioning agent to learn basic thermodynamic response laws, obtain an initial strategy with a safe baseline, and avoid the security risks caused by random exploration from scratch.

[0074] It should be noted that, in order to train the G-QMIX network, this embodiment of the application models the data center air conditioning group control problem as a decentralized partially observable Markov decision process (Dec-POMDP), and its corresponding basic reinforcement learning structure can be as follows: Figure 6 As shown, Figure 6 This is a schematic diagram of the basic reinforcement learning structure in one implementation of the data center air conditioning group control method of this application.

[0075] In the decision-making process of this application embodiment, the state space Designed as the set of all possible states of the environment (such as the states of air conditioners, sensors, electricity meters, etc.), global state It can include the operational metrics of the data center and complete graph structure data, and can be used as input for hybrid networks. That is, during offline reinforcement learning, the state space of reinforcement learning in this embodiment is determined based on the operational metrics of the data center and the graph structure data. Local observation vectors For decentralized execution, each air conditioning agent can be restricted to obtaining only its own operational data and the associated environmental data of the corresponding area.

[0076] ; ; in, This indicates the average temperature and humidity of the computer room, used to reflect the overall thermal environment level; This indicates the total IT load in the data center at the current moment, used to reflect the total cooling demand; It is used to reflect the ambient temperature and can affect the heat exchange efficiency of air conditioners. Indicates air conditioner Power on / off status, Indicates the current set temperature. Indicates the return air vent temperature. This indicates the current operating power. This refers to the associated environmental data for the related regions.

[0077] It is understandable that the aforementioned operational metrics may include global environmental data such as average temperature and humidity in the data center, IT load, and outdoor temperature. The graph structure data described above can be a graph composed of control nodes, environment nodes, and edges.

[0078] In this embodiment of the application, the air conditioning intelligent agent in reinforcement learning The action space can be designed as a discrete set of control commands (such as air conditioner on / off and temperature setting). That is, the action space of each air conditioner is the set of control commands corresponding to that air conditioner.

[0079] In some embodiments of this application, the actions of the air conditioning intelligent agent may include a combination of two dimensions, specifically power on / off actions and temperature setpoints.

[0080] It should be noted that the above power on / off actions are... The aforementioned temperature setpoint is the target return air vent temperature set when the air conditioner is turned on. To facilitate neural network output, this application can flatten the above combination into K discrete action indices (e.g., K=14: action 0 is shutdown, actions 1-13 correspond to different temperature setpoints). The combined action of each air conditioner can be... .

[0081] In this embodiment, the reward function can be designed as a multi-objective composite reward function. This is used to guide the air conditioning agent to achieve low energy consumption while meeting constraints such as temperature and humidity safety. The relationship of the multi-objective composite reward function can be expressed as follows: ; in, Temperature safety is the highest priority reward. This is based on the temperature of any associated sensor. Higher than the preset safety threshold (e.g., 27℃ or other), apply a strong penalty based on the linear distillation function ReLU, as shown below: ; in, This is the temperature penalty weighting coefficient. For sensor indexing.

[0082] It should be noted that humidity safety bonus Humidity for any sensor associated with the air conditioner Beyond the recommended range (For example, when the percentage is 40%-60%), a penalty may be imposed, as shown below: ; in, Humidity penalty weighting coefficient, Indicates the target humidity value. This indicates the allowable humidity range.

[0083] It should be noted that energy efficiency rewards It is negatively correlated with the real-time total power of all air conditioners (related to energy efficiency) and is used to guide energy conservation, as shown below: ; in, This is the energy consumption penalty weighting coefficient. The operating power of the air conditioning intelligent agent, Agent i.

[0084] It should be noted that frequency control penalties To prevent equipment wear caused by frequent start-stop or drastic adjustments of the air conditioner, a penalty is applied to the switching of actions, as shown below: ; in, To control the penalty weighting coefficient. For indicator functions, in hour =1; otherwise, It is 0.

[0085] In practical applications, it can be set Ensuring environmental safety is the top priority.

[0086] It should be noted that a state-action-reward sequence can be constructed based on the states and actions in the offline dataset. This state-action-reward sequence can then be used to pre-train a pre-trained policy model using a behavior cloning / offline reinforcement learning algorithm. The trained model is then evaluated; if it fails to meet the requirements, it is returned for further iteration; if it meets the requirements, it can be used as a pre-trained model for the next stage of simulation environment training. Specifically, the steps of obtaining an offline training dataset and performing reinforcement learning based on it to obtain a pre-trained model include: constructing a state-action-reward sequence based on the state and action data in the offline dataset; training the pre-trained policy model based on the state-action-reward sequence to obtain a trained pre-trained model; and using the trained pre-trained model if it meets the requirements.

[0087] In this embodiment of the application, the state-action pair dataset can be determined based on the offline training dataset. Furthermore, high-quality samples with a Power Usage Effectiveness (PUE) of less than 1.5 and no high-temperature alarms were selected from the state-action pair dataset as the target sample set. The target sample set was then processed using a behavior cloning algorithm, enabling the pre-trained policy model of the air conditioning agent to be optimized. Fitting historical behavior. The training loss function can be cross-entropy loss: ; in, This represents the behavioral cloning loss function. Represents the model parameters of the pre-trained policy model; For the target sample set, This represents high-quality samples in the target sample set. This is the local observation vector (i.e., the state). This refers to the operation of the air conditioner.

[0088] Understandably, when the accuracy of the predicted action compared to the historical action exceeds a preset accuracy threshold (such as 85%, 90%, etc.), the pre-trained model can be considered to have met the standard, and pre-training can be stopped.

[0089] It should be noted that the pre-trained model obtained through the offline training phase already possesses basic PID control logic capabilities and can be safely started. However, because the upper limit of the strategy learned by the model is limited by the optimal strategy in historical data, the system's potential is not fully explored. In this embodiment, a high-fidelity data center simulation environment based on a physics engine (such as EnergyPlus) or mathematical model is constructed during the simulation reinforcement training phase. The pre-trained model from the offline training phase is loaded for large-scale online interactive simulation environment training (Phase Two: Simulation Environment Training). During the simulation reinforcement training phase, high-intensity exploration noise and high hotspot penalty coefficients are introduced to guide the model out of conservative strategies and to explore and master advanced cooperative strategies that utilize "local sacrifice" to achieve "global optimum."

[0090] In some embodiments of this application, a high-fidelity thermal simulation model can be constructed based on the actual physical parameters of the data center (such as size, equipment model, rack layout, etc.), and the model parameters can be calibrated using historical data, with the error controlled within 5%. Based on this, the model parameters of a pre-trained model trained during the offline training phase can be loaded as the initialization parameters of the G-QMIX model. Strategies are explored. The loss can decrease linearly from 0.5 to 0.05. The G-QMIX network parameters can be updated using TD-Error to minimize the loss function. It can be: ; in, For a multi-objective composite reward function, As a reward discount factor, To maximize the expected value of global coordinated actions in the next moment. This represents the value of global coordinated actions at the current moment.

[0091] Understandably, through the simulation-based reinforcement training phase, the model can learn temperature-safe and energy-efficient control strategies. The training results are evaluated by minimizing the loss function. If the completed simulation training model fails to meet the standards, indicating a model requiring improvement, the process can return to the GQMIX simulation training phase, using the model parameters of the model to be improved as initialization parameters. The strategy explores policies and updates the G-QMIX network parameters via TD-Error, thereby achieving policy exploration and optimization. Specifically, the step of performing simulation training based on the pre-trained model to obtain the simulation training model includes: transferring the model parameters of the pre-trained model as initialization parameters to the G-QMIX model; through... The strategy explores the G-QMIX model and updates its parameters using TD-Error to obtain an optimized G-QMIX model; if the optimized G-QMIX is obtained, it is used as the simulation training model.

[0092] It should be noted that, A strategy is a fundamental approach used to balance exploration and exploitation. It randomly selects an action with a probability to explore unknown actions in order to discover possible better strategies.

[0093] Understandably, TD-Error, or TD error, is a metric used to measure the gap between the current global joint action value and the target value. Updating model parameters by minimizing the square of the TD error allows for updating the current value using the current reward and the value estimate for the next time step, enabling single-step updates without waiting for the final result. By continuously reducing the TD error, the estimated global joint action value gradually approaches the true expected cumulative reward, ultimately guiding the air conditioning agent to select the optimal joint action.

[0094] It should be explained that once the simulation training model meets the requirements, it can be loaded into Phase Three: Real Environment Fine-tuning for safety fine-tuning. In this phase, the simulation-trained model can be deployed to a real data center, thereby resolving the "virtual-real difference" issue.

[0095] In some embodiments of this application, the G-QMIX model can be deployed in an edge computing gateway and communicate with the air conditioner via the Modbus / TCP protocol. The G-QMIX model outputs actions. u t Then, a layer of safety constraints for real-world testing can be added: Rule 1: If the temperature T of a sensor associated with a certain cabinet is greater than 28℃ (close to the red line), the air conditioner fan speed in that area must be forced to be set to 100%; Rule 2: The interval between two actions of the same air conditioner must not be less than 5 minutes (to protect the compressor). If the model's output action violates the rules, it can be intercepted and replaced with a safety action.

[0096] It should be noted that online fine-tuning of G-QMIX can be performed based on a real-world environment once the G-QMIX model is deployed. When performing online fine-tuning of G-QMIX, a conservative update strategy can be adopted: setting an extremely low learning rate. Perform small-step updates, such as This avoids disrupting learned strategies; it also allows for experience revisiting: maintaining a small pool of real data experience (Capacity=1000), and updating the gradient descent every 24 hours using the real data collected that day, thereby enabling continuous monitoring and updating of the G-QMIX model after deployment.

[0097] Specifically, the step of fine-tuning the simulation training model to obtain a real environment model includes: collecting data from the real environment to obtain small sample data; updating the simulation training model with a conservative strategy and testing it in the real environment based on the small sample data to obtain a fine-tuned simulation training model; and if the fine-tuned simulation training model meets the deployment criteria, then the fine-tuned simulation training model is used as the real environment model.

[0098] Understandably, the trained real-world model described above can be deployed online and continuously monitored and updated based on experience. The aforementioned small sample data refers to data collected in the deployment data center, such as current running data, current associated environment data, and global environment data.

[0099] It should be understood that during the aforementioned fine-tuning process in the real-world environment, by introducing safety constraints through a safety shield layer, the predicted temperature is validated according to rules before the action is executed, thus intercepting actions with potential overheating risks. Simultaneously, by utilizing a small amount of new data generated in the real environment and employing a very low learning rate to fine-tune the model online, the dynamic deviation between the simulation model and the real physical environment (sim-to-real gap) can be corrected. This allows the model to gradually adapt to disturbances not present in the simulation environment, such as airflow leakage and sensor drift in a real computer room, achieving a smooth implementation of the control strategy.

[0100] This application embodiment obtains an offline training dataset and performs reinforcement learning based on the offline training dataset to obtain a pre-trained model; it then performs simulation training based on the pre-trained model to obtain a simulation-trained model; finally, it fine-tunes the simulation-trained model in a real environment to obtain a real-environment model. By establishing a pre-trained model using the offline training dataset in the first stage, continuing reinforcement learning in the simulation environment to adapt to complex dynamics in the second stage, and performing online fine-tuning in the real environment with a small learning rate in the third stage, combined with parameter freezing, experience replay, and target network mechanisms, the stability and security of the cross-stage transfer process are ensured.

[0101] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the data center air conditioning group control method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0102] This application also provides a group control device for computer room air conditioning; please refer to [reference needed]. Figure 7 , Figure 7 This application provides a schematic diagram of the modular structure of a data center air conditioning group control device, which includes: The data acquisition module 10 is used to acquire the current operating data of different air conditioners in the computer room and the current associated environmental data of the associated area corresponding to each air conditioner; Vector construction module 20 is used to determine the current local observation vector of the target air conditioner based on the current operating data and the current associated environmental data; wherein, the target air conditioner is any air conditioner in the computer room; The local value management module 30 is used to obtain the historical actions of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector. The global control module 40 determines the global combined value of the computer room based on the local action value of each air conditioner, and determines the air conditioning group control scheme of the computer room based on the global combined value.

[0103] The data center air conditioning group control device provided in this application, employing the data center air conditioning group control method in the above embodiments, can solve the technical problem of existing data center air conditioning systems that independently adjust temperature settings and lack inter-device coordination. Compared with the prior art, the beneficial effects of the data center air conditioning group control device provided in this application are the same as those of the data center air conditioning group control method provided in the above embodiments, and other technical features in the data center air conditioning group control device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0104] This application provides a data center air conditioning group control device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the data center air conditioning group control method in the above embodiment 1.

[0105] The following is for reference. Figure 8 The diagram illustrates a structural schematic suitable for implementing the data center air conditioning group control device in the embodiments of this application. The data center air conditioning group control device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), vehicle terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 8 The data center air conditioning group control equipment shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0106] like Figure 8 As shown, the data center air conditioning group control equipment may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the data center air conditioning group control equipment. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, touch screens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the data center air conditioning group control equipment to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows data center air conditioning group control equipment with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0107] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0108] The data center air conditioning group control equipment provided in this application, employing the data center air conditioning group control method in the above embodiments, can solve the technical problem of existing data center air conditioners independently adjusting temperature settings and lacking coordination between devices. Compared with the prior art, the beneficial effects of the data center air conditioning group control equipment provided in this application are the same as those of the data center air conditioning group control method provided in the above embodiments, and other technical features in this data center air conditioning group control equipment are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0109] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0110] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0111] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the computer room air conditioning group control method in the above embodiments.

[0112] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0113] The aforementioned computer-readable storage medium may be included in the computer room air conditioning group control equipment; or it may exist independently and not be installed in the computer room air conditioning group control equipment.

[0114] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the computer room air conditioning group control equipment, cause the computer room air conditioning group control equipment to: Obtain the current operating data of different air conditioners in the computer room and the current associated environmental data of the corresponding associated areas of each air conditioner; For a target air conditioner, the current local observation vector of the target air conditioner is determined based on the current operating data and the current associated environmental data; wherein, the target air conditioner is any air conditioner in the computer room; Obtain the historical actions of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector; The global combined value of the computer room is determined based on the local action value of each air conditioner, and the air conditioning group control scheme of the computer room is determined based on the global combined value.

[0115] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0116] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0117] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0118] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described data center air conditioning group control method. This solves the technical problem of existing data center air conditioning systems that independently adjust temperature settings and lack inter-device coordination. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the data center air conditioning group control method provided in the above embodiments, and will not be repeated here.

[0119] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described data center air conditioning group control method.

[0120] The computer program product provided in this application can solve the technical problem of existing computer room air conditioners that independently adjust temperature settings and lack inter-device coordination. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the computer room air conditioner group control method provided in the above embodiments, and will not be repeated here.

[0121] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the scope of protection of this application.

Claims

1. A method for group control of air conditioning in a computer room, characterized in that, The method includes: Obtain the current operating data of different air conditioners in the computer room and the current associated environmental data of the corresponding associated areas of each air conditioner; For a target air conditioner, the current local observation vector of the target air conditioner is determined based on the current operating data and the current associated environmental data; wherein, the target air conditioner is any air conditioner in the computer room; Obtain the historical actions of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector; The global combined value of the computer room is determined based on the local action value of each air conditioner, and the air conditioning group control scheme of the computer room is determined based on the global combined value.

2. The data center air conditioning group control method as described in claim 1, characterized in that, The step of determining the global combined value of the computer room based on the local action value of each air conditioner includes: Construct a joint input vector based on the local action value of each air conditioner; The joint input vector is corrected by a multi-head self-attention module to obtain a corrected output vector; wherein, the corrected output vector includes the corrected value of each air conditioner, and the corrected value is the value after correcting the local action value; The corrected values ​​of each air conditioner are nonlinearly aggregated based on the topology-aware hybrid network module to obtain the global joint value of the computer room.

3. The data center air conditioning group control method as described in claim 2, characterized in that, The multi-head self-attention module includes: three independent linear transformation units, a multi-head self-attention unit, a residual connection unit, and an output unit; The step of correcting the joint input vector using a multi-head self-attention module to obtain a corrected output vector includes: The joint input vector is input into three independent linear transformation units for linear projection to obtain the feature representation of the joint input vector. The feature representation is input into the multi-head self-attention unit to perform multi-head attention calculation, and attention weights are obtained. The attention weights are residually connected using the residual joint unit to obtain the fused features; The fused features are layer-normalized using the output unit to obtain a corrected output vector.

4. The data center air conditioning group control method as described in claim 2, characterized in that, The topology-aware hybrid network module includes: a graph neural network state encoder, a hybrid network, and a supernetwork corresponding to each layer of the hybrid network; The step of performing nonlinear aggregation of the corrected values ​​of each air conditioner based on the topology-aware hybrid network module to obtain the global joint value of the computer room includes: Obtain the graph structure data corresponding to the computer room, and extract features from the graph structure data through the graph neural network state encoder to obtain a global topology embedding vector; The global topology embedding vector is input into the hypernetwork to determine the hypernetwork weights and hypernetwork biases for each layer of the hypernetwork. The global joint value of the data center is obtained by weighting the modified value using the hybrid network based on the supernetwork weights and the supernetwork biases.

5. The data center air conditioning group control method as described in claim 1, characterized in that, Before the step of obtaining the historical actions of the target air conditioner at the previous moment and determining the current local action value of the target air conditioner based on the historical actions and the current local observation vector, the method further includes: Obtain an offline training dataset and perform reinforcement learning based on the offline training dataset to obtain a pre-trained model; Based on the pre-trained model, simulation training is performed to obtain the simulation training model; The simulation training model is fine-tuned for a real environment to obtain a real environment model; the real environment model includes at least a multi-head self-attention module and a topology-aware hybrid network module.

6. The data center air conditioning group control method as described in claim 5, characterized in that, The state space of the reinforcement learning is determined based on the computer room's operating indicators and graph structure data. The action space of each air conditioner is the set of control commands corresponding to that air conditioner, and the reward function is a multi-objective composite reward function.

7. A group control device for computer room air conditioning, characterized in that, The computer room air conditioning group control device includes: The data acquisition module is used to acquire the current operating data of different air conditioners in the computer room and the current associated environmental data of the associated area corresponding to each air conditioner; A vector construction module is used to determine the current local observation vector of a target air conditioner based on the current operating data and the current associated environmental data; wherein the target air conditioner is any air conditioner in the computer room; The local value management module is used to obtain the historical actions of the target air conditioner at the previous moment, and determine the current local action value of the target air conditioner based on the historical actions and the current local observation vector. The global management module determines the global combined value of the computer room based on the local action value of each air conditioner, and determines the air conditioning group control scheme of the computer room based on the global combined value.

8. A group control device for computer room air conditioning, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the data center air conditioning group control method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the computer room air conditioning group control method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the data center air conditioning group control method as described in any one of claims 1 to 6.

Images