An artificial intelligence-based data machine room energy consumption optimization method and device

Abstract

The application discloses a data machine room energy consumption optimization method and device based on artificial intelligence, and particularly relates to the field of data acquisition monitoring and air conditioner group control, and is used for solving the problem that the existing machine room control mode depends on single-point temperature or experience rule adjustment, and it is difficult to realize global collaborative energy efficiency optimization and energy consumption control of a heat field-load-supply air in a multi-server, multi-air conditioner and multi-airflow coupling environment; the application constructs a high-dimensional state tensor containing a temperature field, load distribution and supply air parameters, establishes a hidden space representation by using a variational autoencoder, trains a hidden space diffusion model to generate multiple candidate future states, screens an optimal future state in combination with an energy efficiency evaluation mechanism, and reconstructs air conditioner supply air parameters, so that dynamic optimization control of machine room energy consumption is realized, operation stability and energy utilization efficiency are improved, response capability to abnormal working conditions and overall collaborative regulation and control level are enhanced, and fine management and intelligent operation and maintenance capability of a data machine room are improved.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and monitoring and air conditioning control technology, and more specifically, to a method and device for optimizing energy consumption in a data center based on artificial intelligence. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence applications, data centers, as core infrastructure for information processing and storage, are experiencing continuous growth in operational scale and power consumption. Data centers typically house numerous server racks, forming a spatial structure with alternating cold and hot aisles. Servers generate significant heat during high-load operation; if heat dissipation is inadequate, hotspots can easily form in localized areas, affecting equipment stability and shortening its lifespan. To ensure the safe operation of servers, data centers are usually equipped with multiple precision air conditioners forming a group control system, which regulates the temperature field by adjusting parameters such as supply air temperature, air volume, and fan speed.

[0003] However, the thermal field distribution inside the data center is highly coupled with the server load status, and the airflow organization is significantly affected by the spatial structure. Load fluctuations, business migration, or changes in air conditioning parameters may all cause dynamic changes in the temperature field. Traditional control methods based on empirical rules or single index feedback are difficult to accurately characterize the spatial relationship between the three-dimensional temperature field and the load distribution. They can often only adjust based on the average temperature or single-point temperature, which can easily lead to local overcooling or overheating, resulting in low energy utilization efficiency.

[0004] Meanwhile, when there is an abnormal surge in load or the air conditioning deviates from the baseline state, the existing control strategies usually lack the ability to predict the future thermal field evolution trend, making it difficult to comprehensively evaluate and optimize among multiple possible operating states, which limits the further improvement of the overall energy consumption optimization and refined control level of the data center. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a data center energy consumption optimization method and equipment based on artificial intelligence to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An artificial intelligence-based method for optimizing energy consumption in data centers includes the following steps:

[0008] S1. Obtain the three-dimensional temperature field of the entire data center, the load distribution of each server node, and the air conditioning supply parameters at the set continuous monitoring time, and construct a high-dimensional state tensor.

[0009] S2. The high-dimensional state tensor of two adjacent historical continuous monitoring moments is compressed into a low-dimensional latent space vector by a variational autoencoder, and a latent space state transition sample set is established.

[0010] S3. Perform a forward-noise Markov chain on the latent space vectors in the state transition sample set, and train the latent space diffusion model by adjusting the network parameters of the inverse denoising network.

[0011] S4. When an abnormal signal of the power environment of the computer room is detected, obtain the current hidden space vector corresponding to the current moment, input the current hidden space vector into the hidden space diffusion model to iteratively perform noise prediction and removal, and generate multiple candidate future hidden space vectors.

[0012] S5. Evaluate the energy efficiency of each candidate future latent space vector and select the target future high-dimensional state tensor with the highest energy efficiency evaluation value.

[0013] S6. Based on the target future high-dimensional state tensor with the highest energy efficiency evaluation value, reconstruct the air supply parameters of the air conditioner.

[0014] As a further aspect of the present invention, in S1, constructing the high-dimensional state tensor specifically includes:

[0015] The air supply temperature setpoint and air supply volume setpoint are collected from the air conditioning group control system and used as air conditioning air supply parameters;

[0016] Time-series temperature data are collected from preset temperature monitoring points in the cold and hot aisles and cabinet air inlets of the computer room, and a three-dimensional temperature field matrix covering the entire computer room is generated through spatial interpolation.

[0017] Collect time-series data on CPU utilization and memory usage of each server, and map them into a load distribution density matrix based on the server's physical space coordinates.

[0018] The three-dimensional temperature field matrix and load distribution density matrix are aligned and stitched together point by point in the spatial dimension. The air conditioning supply parameters are added to construct a high-dimensional state tensor containing the temperature field, load distribution and air conditioning supply parameters, and stored in the historical monitoring database.

[0019] As a further aspect of the present invention, in step S2, establishing the latent space state transition sample set specifically includes:

[0020] Read the high-dimensional state tensors of two consecutive adjacent monitoring times from the historical monitoring database to form a time-series state pair;

[0021] The two high-dimensional state tensors in the temporal state pair are respectively input into the encoding network of the variational autoencoder. The encoding network consists of three-dimensional convolutional layers and fully connected layers connected in sequence. By setting the stride of the convolutional layer to be greater than 1, the spatial resolution is downsampled to compress the input tensor into a feature vector in the latent space.

[0022] Based on the feature vectors of the latent space, the mean vector and log-variance vector of the latent space distribution are generated in the fully connected layer, and the latent space vector is generated by reparameterized sampling.

[0023] The latent space vectors generated from all time-series state pairs in the historical database are paired with the latent space vectors of the previous and next time steps to form latent space state transition pairs, and then summarized to form a latent space state transition sample set.

[0024] As a further aspect of the present invention, in S3, training the latent space diffusion model specifically includes:

[0025] The latent space state transition sample set is trained in batches. Each training sample corresponds to a latent space state transition pair, which includes the latent space vector of the previous time step and the latent space vector of the next time step.

[0026] Gaussian noise is progressively superimposed on the latent space vector at the next time step according to a preset noise schedule. The noise component added at each time step is used as a supervision label to generate the noisy latent space vector corresponding to each time step.

[0027] The noisy latent space vector, time step index, and previous time step latent space vector corresponding to the sample at each time step are concatenated and input into the inverse denoising network, which outputs the predicted noise component. The inverse denoising network consists of a time step embedding layer, a three-dimensional convolutional layer, and a skip connection.

[0028] The error between the predicted noise component and the supervision label is calculated. The convolution kernel weights and bias terms of the inverse denoising network are adjusted through backpropagation, and the inverse denoising network is preserved as a latent space diffusion model.

[0029] As a further aspect of the present invention, in step S4, the abnormal signal of the computer room power environment is generated when the temperature field distribution gradient, load fluctuation amplitude, or air conditioning supply parameters deviate from their respective preset dynamic baselines.

[0030] As a further aspect of the present invention, in step S4, generating multiple candidate future latent space vectors specifically includes:

[0031] When an abnormal signal of the power environment of the computer room is received, the high-dimensional state tensor of the current moment is obtained and input into the variational autoencoder network to obtain the current latent space vector.

[0032] A random noise latent space vector that follows a standard normal distribution is generated as the initial state for inverse diffusion. The number of inverse diffusion steps is set, and the initial state for inverse diffusion and the current latent space vector are jointly input into the latent space diffusion model. The noise component of the current step is predicted and subtracted one by one, and multiple restored candidate future latent space vectors are output.

[0033] As a further aspect of the present invention, in step S5, the selection of the target future high-dimensional state tensor with the highest energy efficiency evaluation value specifically includes:

[0034] Each candidate future latent space vector is input into the decoding network of the variational autoencoder. The decoding network consists of sequentially connected deconvolutional layers. By setting the stride of the deconvolutional layers to be greater than 1, the spatial resolution is upsampled to reconstruct the corresponding candidate future high-dimensional state tensor.

[0035] Extract the temperature field matrix, load distribution density matrix and air conditioning supply parameters from the candidate future high-dimensional state tensor. Calculate the correlation coefficient between the load distribution density matrix and the temperature field matrix at the corresponding spatial locations. At the same time, calculate the absolute value of the difference between the global mean of the temperature field matrix and the supply temperature set value in the air conditioning supply parameters as a deviation index.

[0036] Energy efficiency evaluation values ​​are constructed by weighting the correlation coefficient as the positive contribution and the deviation index as the penalty.

[0037] The energy efficiency evaluation values ​​of all candidate future high-dimensional state tensors are compared, and the candidate future high-dimensional state tensor with the highest energy efficiency evaluation value is determined as the target future high-dimensional state tensor.

[0038] As a further aspect of the present invention, in step S6, reconstructing the air conditioning supply parameters based on the target future high-dimensional state tensor with the highest energy efficiency evaluation value specifically includes:

[0039] The air conditioning supply parameters are extracted from the future high-dimensional state tensor. The corresponding supply temperature setpoint is converted into the temperature setting command of the air conditioning group control system. The supply air volume setpoint is converted into the fan speed adjustment command. The temperature setting command and the fan speed adjustment command are sent to the air conditioning group control system for execution through the industrial control bus.

[0040] On the other hand, the present invention provides an artificial intelligence-based data center energy consumption optimization device, comprising:

[0041] The data acquisition and monitoring host is used for the processing and centralized management of temperature, server and air conditioning data in the data center, and also for receiving control commands and transmitting corresponding adjustment commands to the air conditioning group.

[0042] Data center display terminals are used to visualize the current operating status of the computer room and the data corresponding to the candidate future high-dimensional state tensors.

[0043] Air conditioning groups are used to form a group control system for air conditioning in a computer room. They receive the setpoint for air supply temperature and the air supply volume adjustment command issued by the data acquisition and monitoring host, and coordinate the control of each air conditioning unit to adjust the air supply status of the computer room and match it with changes in server load.

[0044] The present invention discloses a data center energy consumption optimization method and equipment based on artificial intelligence, highlighting its technical effects and advantages.

[0045] This invention achieves multi-path generation and evaluation of the future operating state of the data center by uniformly modeling the three-dimensional temperature field, server load distribution, and air conditioning supply parameters of the entire data center, constructing a high-dimensional state tensor, and introducing a latent space modeling and diffusion prediction mechanism. At the same time, it evaluates the energy efficiency of multiple candidate future states and selects the optimal target state, and then reconstructs the air conditioning supply parameters accordingly to achieve optimized control decisions based on future prediction results. This avoids the local optimum problem caused by a single prediction, reduces excessive cooling and ineffective air supply while meeting temperature control safety constraints, improves energy utilization efficiency, and reduces overall operating energy consumption.

[0046] It achieves global collaborative energy efficiency optimization of thermal field, load and air supply in multi-server, multi-air conditioner and multi-airflow coupling environment, and performs multi-path evolution prediction and optimal control decision under abnormal conditions, promoting the development of data center towards refinement, intelligence and energy saving. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of an artificial intelligence-based energy consumption optimization method for data centers according to the present invention;

[0048] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based data center energy consumption optimization device according to the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] Figure 1 This invention presents a data center energy consumption optimization method based on artificial intelligence, which includes the following steps:

[0052] S1. Obtain the three-dimensional temperature field of the entire data center, the load distribution of each server node, and the air conditioning supply parameters at the set continuous monitoring time, and construct a high-dimensional state tensor.

[0053] S2. The high-dimensional state tensor of two adjacent historical continuous monitoring moments is compressed into a low-dimensional latent space vector by a variational autoencoder, and a latent space state transition sample set is established.

[0054] S3. Perform a forward-noise Markov chain on the latent space vectors in the state transition sample set, and train the latent space diffusion model by adjusting the network parameters of the inverse denoising network.

[0055] S4. When an abnormal signal of the power environment of the computer room is detected, obtain the current hidden space vector corresponding to the current moment, input the current hidden space vector into the hidden space diffusion model to iteratively perform noise prediction and removal, and generate multiple candidate future hidden space vectors.

[0056] S5. Evaluate the energy efficiency of each candidate future latent space vector and select the target future high-dimensional state tensor with the highest energy efficiency evaluation value.

[0057] S6. Based on the target future high-dimensional state tensor with the highest energy efficiency evaluation value, reconstruct the air supply parameters of the air conditioner.

[0058] In S1, a high-dimensional state tensor is constructed.

[0059] The air conditioning group control system comprises multiple precision air conditioning units, each with independent supply air temperature setpoints and supply air volume adjustment values. The system periodically reads the currently effective supply air temperature setpoints and corresponding fan speeds or air volume percentages of each unit via an industrial control bus, and synchronizes them with a unified timestamp. The data acquisition cycle is set to once every 30 seconds, determined after analyzing the thermal inertia characteristics of the computer room to ensure a recognizable response relationship between temperature changes and control changes. The supply air parameters of multiple air conditioning units are numbered according to their spatial arrangement in the computer room and formed into a vectorized structure. The supply air temperature setpoints and air volume adjustment values ​​of each unit are arranged in a fixed order to form an air conditioning supply air parameter set. During the acquisition process, if a communication interruption or abnormal parameter jump is detected, the valid value of the previous cycle is retained and an anomaly is recorded to ensure data continuity and time consistency, thereby obtaining a stable and traceable sequence of air conditioning supply air parameter data.

[0060] Temperature monitoring points were deployed in the cold aisle, hot aisle, and air inlets of each server rack in the computer room. These monitoring points were layered according to the rack row spacing and height; for example, sensors were placed on the top, middle, and bottom layers of each rack row. Auxiliary monitoring points were added in the middle of the cold and hot aisles to improve spatial resolution. Time-series temperature data from all monitoring points were collected, and a spatial coordinate index table was created based on the physical spatial coordinates of each monitoring point. A regular three-dimensional grid was constructed based on the computer room's floor plan dimensions and floor height information. Spatial interpolation was performed on the temperature values ​​of the discrete monitoring points using an inverse distance-weighted method, with distance weights calculated based on actual physical distances. The interpolated values ​​resulted in a three-dimensional temperature field matrix that corresponded one-to-one with the computer room's spatial grid. Each grid point in the matrix corresponds to a specific spatial coordinate and temperature value.

[0061] The CPU utilization and memory usage data of each server are periodically acquired through the server operation monitoring interface in the data center and bound to the server rack number and installation location. Each server in the data center has fixed physical coordinate information, including its rack number, floor level, and corresponding spatial location. CPU utilization and memory usage are weighted and fused according to a set ratio to form a comprehensive load index. This ratio is determined using historical power consumption test data; for example, CPU utilization is weighted at 0.7 and memory usage at 0.3 to reflect the differences in heat contribution from different resources. Subsequently, based on the server's spatial coordinates, the comprehensive load index is mapped to a three-dimensional grid structure similar to the temperature field. If multiple servers reside in the same grid, their comprehensive load indices are accumulated to form a load distribution density matrix, ensuring that each spatial grid corresponds to a specific load intensity value.

[0062] The spatial coordinate consistency of the three-dimensional temperature field matrix and the load distribution density matrix is ​​verified to ensure that the number of grids, resolution, and origin of coordinates are consistent. For cases with resolution differences, resampling is performed according to a unified grid standard to ensure point-to-point correspondence between temperature and load data in the spatial dimension. Subsequently, temperature and load values ​​are concatenated along the channel dimension according to the spatial grid order to form a spatial data structure containing multi-channel features. The air conditioning supply parameter vector for the current period is appended to the end of this structure to form a high-dimensional state tensor containing both spatial and control features. Each state tensor is accompanied by a timestamp and anomaly marker information, and is written to the historical monitoring database in chronological order, with storage capacity managed using a cyclic overwrite method.

[0063] In S2, a hidden space state transition sample set is established.

[0064] The historical monitoring database stores continuously acquired high-dimensional state tensors in chronological order. Each state tensor contains a timestamp and a corresponding three-dimensional spatial data structure. During reading, the data is sorted from earliest to latest according to the timestamp and filtered at fixed time intervals, such as extracting two consecutive records at 30-second intervals to ensure there are no time jumps between adjacent high-dimensional state tensors. Records with missing data or anomaly markers are directly removed from that time period to maintain the continuity of the time-series data. The two high-dimensional state tensors read retain their original spatial resolution and channel structure. The state tensor at the previous moment is recorded as the previous state data, and the state tensor at the next moment is recorded as the subsequent state data. They are then paired in chronological order to form a unified and continuous time-series state pair.

[0065] The encoding network employs a 3D convolutional structure, with input being a high-dimensional state tensor containing channels for temperature field, load distribution, and air supply parameters. The network structure comprises three 3D convolutional layers, each with a kernel size of 3×3×3 and a stride of 2 (or another stride greater than 1), achieving a resolution reduction of half in the spatial dimension through downsampling. Each convolutional operation is followed by a batch normalization layer and a non-linear activation layer to enhance feature representation. After three consecutive convolutional downsampling layers, the original 3D spatial resolution is compressed to one-eighth of its original value, and spatial features are aggregated into a higher-level semantic representation. The multi-channel feature map output from the convolution is then unfolded into a one-dimensional feature vector and input to a fully connected layer for further feature integration. The number of neurons in the fully connected layer is set to 256, a value determined by comparing historical data compression and reconstruction errors. After this processing, the input high-dimensional state tensor is compressed into a fixed-length latent space feature vector. In the final stage of the encoding network, these latent space feature vectors are input to two independent fully connected branch structures. The first branch outputs the mean vector of the latent space distribution, and the second branch outputs the corresponding log-variance vector. Both have the same length as the latent space dimension; in this embodiment, the latent space dimension is set to 128, a value determined by balancing model reconstruction accuracy and computational complexity. After generating the mean and log-variance vectors, a reparameterization method is used for sampling. First, a standard normal random vector with the same latent space dimension is generated. Then, the scales corresponding to this random vector and the log-variance vector are transformed element-wise, and finally, they are element-wise superimposed with the mean vector to obtain the final latent space vector. This process ensures that the sampling operation can participate in backpropagation training and also gives the latent space representation continuous probability distribution characteristics, thereby achieving smooth modeling of the data center's operating status.

[0066] The above encoding and sampling operations are performed sequentially on all temporal state pairs in the historical database to obtain the corresponding latent space vectors for the previous and next time steps. Each pair of latent space vectors is then grouped into a state transition pair according to chronological order, where the previous latent space vector represents the state's starting point, and the next latent space vector represents the state's ending point. All state transition pairs are aggregated and stored in chronological order to form a latent space state transition sample set, retaining the corresponding time interval information for subsequent diffusion model training. Sample pairs with abnormal labels or inconsistent time intervals are removed to ensure stable data distribution and temporal continuity within the sample set, thereby constructing a complete latent space state transition training data structure.

[0067] In S3, the latent space diffusion model is trained.

[0068] The latent space state transition sample set consists of a large number of state transition pairs, each containing the latent space vector from the previous time step and the latent space vector from the next time step. During training, the samples are processed in batches with a fixed batch size, for example, 64 state transition pairs are selected per batch. The batch size is determined after testing the memory capacity and training stability. For the latent space vector from the next time step in each sample, a progressive noise addition operation is performed according to a predefined noise schedule. The noise schedule defines the noise intensity at different time steps. In this embodiment, the total number of diffusion steps is set to 1000 steps, and the noise intensity gradually increases from a smaller value to a larger value. The specific values ​​are determined by statistically analyzing the distribution of the latent space vectors; for example, the initial step noise standard deviation is set to 0.0001, and the final step to 0.02, arranged in a linearly increasing manner. Gaussian noise of corresponding intensity is superimposed on the latent space vector from the next time step at each time step. The noise vector has the same dimension as the latent space, and the superimposed noise component is recorded as a supervision label at each step. In this way, a series of noisy latent space vectors can be obtained under different levels of noise degradation, enabling the model to learn the ability to recover the true latent space vector from arbitrary noise levels. The entire noise-adding process is repeated for each sample, forming a training data set covering all time steps.

[0069] At each training time step, the corresponding noisy latent space vector is concatenated with the previous time step's latent space vector of the sample along the channel dimension. Simultaneously, the current time step index is input into the time step embedding layer. The time step embedding layer converts integer time step indices into fixed-length embedding vectors; in this embodiment, the embedding dimension is set to 128, generated through a preset lookup table mapping method. The embedding vector and the concatenated latent space features serve as input to the inverse denoising network. The inverse denoising network employs a three-dimensional convolutional structure, containing several convolutional blocks. Each convolutional block consists of a three-dimensional convolutional layer, a normalization layer, and a non-linear activation layer, establishing skip connections between the encoding and decoding paths to fuse high-level semantic features with low-level detail features. The network outputs a predicted noise component with the same dimension as the noisy latent space vector. Subsequently, the mean square error between the predicted noise component and the recorded true noise component is calculated. The error value is backpropagated to each layer of the network through an automatic differentiation mechanism, updating the convolutional kernel weights and bias terms. Training uses a fixed learning rate, for example, 0.0001. The optimization algorithm is an adaptive gradient method, and the number of training iterations is set to 100 rounds. After each training round, the overall error change is statistically analyzed until the error converges and tends to stabilize. After training is completed, the final parameters are fixed and the current network structure and weights are saved to form a latent space diffusion model.

[0070] In step S4, multiple candidate future latent space vectors are generated.

[0071] During the operation of the data center, a high-dimensional state tensor is continuously generated according to a fixed sampling period. When an anomaly signal is triggered, the high-dimensional state tensor corresponding to the trigger moment is immediately locked. This high-dimensional state tensor maintains the same spatial resolution and channel structure as in the historical training phase, including channels for 3D temperature field data, load distribution density, and air conditioning supply parameters. This high-dimensional state tensor is then input into a variational autoencoder network that has been trained and has fixed parameters. The encoding network structure remains consistent with the training phase, with each convolutional layer having a stride greater than 1 to achieve spatial compression. The input tensor first undergoes convolutional feature extraction and spatial downsampling, then is mapped to a latent space feature representation through a fully connected layer, and the current latent space vector is generated through reparameterized sampling of the mean vector and log-variance vector. During normal operation of the data center, the gradient changes of the 3D temperature field in space are continuously calculated. The gradient value is obtained by dividing the temperature difference between adjacent spatial grids by the actual physical distance, and the global average gradient and the maximum local gradient are statistically analyzed. When the average gradient or local gradient exceeds the dynamic baseline interval obtained from the historical stable operation phase, an anomaly marker is generated. The dynamic baseline is obtained by statistically analyzing the operating data of the most recent three to seven days using a sliding window. For example, a normal fluctuation range is defined as plus or minus two standard deviations of the historical average. For server load fluctuations, the change in the comprehensive load index between two consecutive sampling periods is calculated. An anomaly detection is triggered when the change exceeds three times the historical average fluctuation value. For air conditioning supply parameters, the deviation between the current setpoint and the historical stable range average is compared. An anomaly signal is generated when the supply air temperature deviates by more than 2 degrees Celsius or the airflow adjustment value deviates by more than 15% of the rated value. If any of the above conditions are met, an abnormal signal for the data center's power environment is output, ensuring that anomaly triggering has a clear quantitative basis.

[0072] After an abnormal signal is triggered and the current latent space vector is obtained, a standard normal random vector with the same dimensions as the latent space is first generated as the initial state for inverse diffusion. Each dimension of this random vector independently follows a normal distribution with a mean of 0 and a standard deviation of 1. The number of inverse diffusion steps is consistent with the training phase, for example, set to 1000 steps. In each step of the inverse diffusion process, the noise latent vector of the current step, the index embedding vector of the corresponding time step, and the current latent space vector are concatenated and input into the trained latent space diffusion model. The model outputs the predicted noise component of the current step and subtracts the predicted noise element-wise from the current noise latent vector. At the same time, the latent vector state is updated in combination with the scale factor of the corresponding time step in the noise scheduling table. This process is executed sequentially from large to small time steps until all inverse diffusion steps are completed, finally obtaining a restored future latent space vector. By changing the initial random noise vector and repeating the above inverse diffusion process, multiple different candidate future latent space vectors can be generated, reflecting multiple future evolution paths that may occur under the current abnormal state conditions.

[0073] In step S5, the target future high-dimensional state tensor with the highest energy efficiency evaluation value is selected.

[0074] The decoding network structure of the variational autoencoder corresponds to that of the encoding network, employing multiple layers of three-dimensional deconvolution to progressively restore spatial resolution. Each candidate future latent space vector is first input into a fully connected layer for dimensional expansion, mapping the one-dimensional latent space representation to a three-dimensional feature block with a fixed number of channels and low spatial resolution, for example, an 8×8×8 spatial size with multiple feature channels. Subsequently, spatial upsampling is performed through three layers of three-dimensional deconvolution, with a stride of 2 for each deconvolution layer, progressively expanding the spatial size to 16×16×16, 32×32×32 until the spatial resolution is restored to be consistent with the original high-dimensional state tensor. After each deconvolution layer, a normalization layer and a nonlinear activation layer are connected to ensure a smooth feature restoration process. The final output multi-channel three-dimensional tensor corresponds to the temperature field data channel, load distribution channel, and air conditioning supply parameter channel according to a pre-defined channel order, thereby reconstructing the candidate future high-dimensional state tensor. This tensor structure is consistent with historical monitoring data.

[0075] The candidate future high-dimensional state tensors obtained from decoding are separated by channel to extract the three-dimensional temperature field matrix, load distribution density matrix, and air conditioning supply parameter vector. For the temperature field matrix and load distribution density matrix, they are mapped point-by-point on a unified spatial grid, and the correlation coefficient between them at the same spatial location is calculated. The correlation coefficient is calculated using a standardized correlation method. First, the temperature and load values ​​are normalized to their mean. Then, the average of the products at corresponding locations is calculated and divided by the standard deviation product to obtain a correlation value ranging from -1 to 1, reflecting the consistency between the cooling distribution and the load distribution. Simultaneously, the global mean of the temperature field matrix is ​​calculated and compared with the supply air temperature setpoint in the air conditioning supply parameters. The absolute value of the difference between the two is taken as the deviation index. The deviation index reflects the degree of matching between the current cooling strategy and the target supply air setting. The correlation coefficient is chosen to characterize the spatial matching degree between temperature distribution and load distribution. When the overlap between the cooling area and the high-load area is high, it indicates that the cooling capacity is concentrated in the heat-generating area, resulting in higher cooling efficiency. The deviation index is chosen to measure the deviation between the average temperature across the entire area and the air supply setpoint. A smaller deviation indicates stable cooling control and no over-cooling. Both reflect the rationality of cooling capacity utilization and the stability of temperature control, respectively, and therefore can be used as a basis for energy efficiency evaluation.

[0076] The correlation coefficient and deviation index are normalized to ensure they fall within a unified numerical range. The normalization range is determined through statistical analysis of historical data; for example, the correlation coefficient is linearly mapped to the 0-1 range based on its historical minimum and maximum values, and the deviation index is scaled proportionally to its historical maximum allowable deviation. These are then combined using a fixed weighting ratio, determined through multiple rounds of comparative experiments on historical energy consumption data and temperature control stability data. The weighted result is used as the energy efficiency evaluation value, and the energy efficiency evaluation values ​​of all candidate future high-dimensional state tensors are ranked. The candidate future high-dimensional state tensor with the highest value is selected as the target future high-dimensional state tensor. The spatial temperature distribution and air supply parameters corresponding to this target state are determined as the optimal evolution result.

[0077] In step S6, the air conditioning supply parameters are reconstructed based on the target future high-dimensional state tensor with the highest energy efficiency evaluation value.

[0078] The target high-dimensional state tensor stores temperature field data, load distribution data, and air conditioning supply parameter data in a fixed channel order. First, the air conditioning supply parameter vector is directly extracted based on the channel index position. This vector contains the supply temperature setpoint and supply air volume setpoint for each air conditioning unit. For the supply temperature setpoint, data conversion is performed according to the control protocol format used by the air conditioning group control system. For example, floating-point temperature values ​​are converted to fixed-point numbers with one decimal place and mapped to the corresponding register address. If the allowable temperature adjustment range of the air conditioning equipment is 16℃ to 30℃, the temperature setpoint is validated before conversion to ensure the output value is within the allowable range; values ​​outside the range are truncated according to boundary values. For the supply air volume setpoint, it is proportionally converted as a percentage of the equipment's rated air volume. For example, the predicted 0.75 is represented as 75% of the rated air volume. Then, based on the calibration curve between fan speed and air volume, the air volume percentage is converted into a specific fan speed value. The calibration curve is determined using factory test data, such as the 100% airflow corresponding to a rated frequency of 50Hz, which is then proportionally converted to obtain the specific frequency value. After conversion, structured temperature setting commands and fan speed adjustment commands are generated as data frames.

[0079] Control commands are issued via an industrial control bus, with the specific protocol fixed during system deployment. Temperature setting commands and fan speed adjustment commands for each air conditioning unit are encapsulated according to the device address, generating a complete control message containing the device address code, function code, data length, and data fields. After generation, the message is sent sequentially to each air conditioning unit controller in polling order, and a confirmation response code is awaited after transmission. If a normal response is received, the control status of that unit is recorded as successful; if no response is received or the response is abnormal, the command is resent once, and an error flag is recorded. All command issuance is completed within one control cycle, for example, updating all unit commands within a 30-second control cycle. After command execution, the air conditioning group control system adjusts the actual operating status based on the updated temperature setpoint and fan speed values, achieving synchronous adjustment of the air supply conditions in the computer room. In this way, the optimized air supply parameters in the future high-dimensional state tensor are accurately mapped to physically executable control actions, completing the conversion from predicted results to actual control behavior.

[0080] Example 2

[0081] The difference between Embodiment 2 and Embodiment 1 is that this embodiment introduces an energy consumption optimization device for data centers based on artificial intelligence.

[0082] Figure 2 A schematic diagram of an AI-based data center energy consumption optimization device is provided. The AI-based data center energy consumption optimization device includes:

[0083] The data acquisition and monitoring host is used for the processing and centralized management of temperature, server and air conditioning data in the data center, and also for receiving control commands and transmitting corresponding adjustment commands to the air conditioning group.

[0084] Data center display terminals are used to visualize the current operating status of the computer room and the data corresponding to the candidate future high-dimensional state tensors.

[0085] Air conditioning groups are used to form a group control system for air conditioning in a computer room. They receive the setpoint for air supply temperature and the air supply volume adjustment command issued by the data acquisition and monitoring host, and coordinate the control of each air conditioning unit to adjust the air supply status of the computer room and match it with changes in server load.

[0086] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0087] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0088] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0089] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0090] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0091] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0092] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0093] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] 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.

[0095] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data center energy consumption optimization method based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the three-dimensional temperature field of the entire data center, the load distribution of each server node, and the air conditioning supply parameters at the set continuous monitoring time, and construct a high-dimensional state tensor. S2. The high-dimensional state tensor of two adjacent historical continuous monitoring moments is compressed into a low-dimensional latent space vector by a variational autoencoder, and a latent space state transition sample set is established. S3. Perform a forward-noise Markov chain on the latent space vectors in the state transition sample set, and train the latent space diffusion model by adjusting the network parameters of the inverse denoising network. S4. When an abnormal signal of the power environment of the computer room is detected, obtain the current hidden space vector corresponding to the current moment, input the current hidden space vector into the hidden space diffusion model to iteratively perform noise prediction and removal, and generate multiple candidate future hidden space vectors. S5. Evaluate the energy efficiency of each candidate future latent space vector and select the target future high-dimensional state tensor with the highest energy efficiency evaluation value. S6. Based on the target future high-dimensional state tensor with the highest energy efficiency evaluation value, reconstruct the air supply parameters of the air conditioner.

2. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In S1, constructing the high-dimensional state tensor specifically includes: The air supply temperature setpoint and air supply volume setpoint are collected from the air conditioning group control system and used as air conditioning air supply parameters; Time-series temperature data are collected from preset temperature monitoring points in the cold and hot aisles and cabinet air inlets of the computer room, and a three-dimensional temperature field matrix covering the entire computer room is generated through spatial interpolation. Collect time-series data on CPU utilization and memory usage of each server, and map them into a load distribution density matrix based on the server's physical space coordinates. The three-dimensional temperature field matrix and load distribution density matrix are aligned and stitched together point by point in the spatial dimension. The air conditioning supply parameters are added to construct a high-dimensional state tensor containing the temperature field, load distribution and air conditioning supply parameters, and stored in the historical monitoring database.

3. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In S2, establishing the latent space state transition sample set specifically includes: Read the high-dimensional state tensors of two consecutive adjacent monitoring times from the historical monitoring database to form a time-series state pair; The two high-dimensional state tensors in the temporal state pair are respectively input into the encoding network of the variational autoencoder. The encoding network consists of three-dimensional convolutional layers and fully connected layers connected in sequence. By setting the stride of the convolutional layer to be greater than 1, the spatial resolution is downsampled to compress the input tensor into a feature vector in the latent space. Based on the feature vectors of the latent space, the mean vector and log-variance vector of the latent space distribution are generated in the fully connected layer, and the latent space vector is generated by reparameterized sampling. The latent space vectors generated from all time-series state pairs in the historical database are paired with the latent space vectors of the previous and next time steps to form latent space state transition pairs, and then summarized to form a latent space state transition sample set.

4. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In S3, training the latent space diffusion model specifically includes: The latent space state transition sample set is trained in batches. Each training sample corresponds to a latent space state transition pair, which includes the latent space vector of the previous time step and the latent space vector of the next time step. Gaussian noise is progressively superimposed on the latent space vector at the next time step according to a preset noise schedule. The noise component added at each time step is used as a supervision label to generate the noisy latent space vector corresponding to each time step. The noisy latent space vector, time step index, and previous time step latent space vector corresponding to the sample at each time step are concatenated and input into the inverse denoising network, which outputs the predicted noise component. The inverse denoising network consists of a time step embedding layer, a three-dimensional convolutional layer, and a skip connection. The error between the predicted noise component and the supervision label is calculated. The convolution kernel weights and bias terms of the inverse denoising network are adjusted through backpropagation, and the inverse denoising network is preserved as a latent space diffusion model.

5. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In S4, the abnormal signal of the computer room power environment is generated when the temperature field distribution gradient, load fluctuation amplitude or air conditioning supply parameters deviate from their respective preset dynamic baselines.

6. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In step S4, generating multiple candidate future latent space vectors specifically includes: When an abnormal signal of the power environment of the computer room is received, the high-dimensional state tensor of the current moment is obtained and input into the variational autoencoder network to obtain the current latent space vector. A random noise latent space vector that follows a standard normal distribution is generated as the initial state for inverse diffusion. The number of inverse diffusion steps is set, and the initial state for inverse diffusion and the current latent space vector are jointly input into the latent space diffusion model. The noise component of the current step is predicted and subtracted one by one, and multiple restored candidate future latent space vectors are output.

7. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In step S5, the specific steps for selecting the target future high-dimensional state tensor with the highest energy efficiency evaluation value include: Each candidate future latent space vector is input into the decoding network of the variational autoencoder. The decoding network consists of sequentially connected deconvolutional layers. By setting the stride of the deconvolutional layers to be greater than 1, the spatial resolution is upsampled to reconstruct the corresponding candidate future high-dimensional state tensor. Extract the temperature field matrix, load distribution density matrix and air conditioning supply parameters from the candidate future high-dimensional state tensor. Calculate the correlation coefficient between the load distribution density matrix and the temperature field matrix at the corresponding spatial locations. At the same time, calculate the absolute value of the difference between the global mean of the temperature field matrix and the supply temperature set value in the air conditioning supply parameters as a deviation index. Energy efficiency evaluation values ​​are constructed by weighting the correlation coefficient as the positive contribution and the deviation index as the penalty. The energy efficiency evaluation values ​​of all candidate future high-dimensional state tensors are compared, and the candidate future high-dimensional state tensor with the highest energy efficiency evaluation value is determined as the target future high-dimensional state tensor.

8. The data center energy consumption optimization method based on artificial intelligence according to claim 1, characterized in that, In step S6, the reconstruction of the air conditioning supply parameters based on the target future high-dimensional state tensor with the highest energy efficiency evaluation value specifically includes: The air conditioning supply parameters are extracted from the future high-dimensional state tensor. The corresponding supply temperature setpoint is converted into the temperature setting command of the air conditioning group control system. The supply air volume setpoint is converted into the fan speed adjustment command. The temperature setting command and the fan speed adjustment command are sent to the air conditioning group control system for execution through the industrial control bus.

9. An AI-based data center energy consumption optimization device, used to implement the AI-based data center energy consumption optimization method according to any one of claims 1-8, characterized in that, include: The data acquisition and monitoring host is used for the processing and centralized management of temperature, server and air conditioning data in the data center, and also for receiving control commands and transmitting corresponding adjustment commands to the air conditioning group. Data center display terminals are used to visualize the current operating status of the computer room and the data corresponding to the candidate future high-dimensional state tensors. Air conditioning groups are used to form a group control system for air conditioning in a computer room. They receive the setpoint for air supply temperature and the air supply volume adjustment command issued by the data acquisition and monitoring host, and coordinate the control of each air conditioning unit to adjust the air supply status of the computer room and match it with changes in server load.

Images