A method for data center chiller system prediction and shutdown temperature setting optimization

Abstract

The application discloses a kind of data center refrigeration system prediction and shutdown temperature setting optimization method, comprising: obtaining the time series data when refrigeration system runs;Time series segmentation clustering model is used to identify and classify the time series data obtained in stage using manual labeling method, and sample data set is constructed;The start-stop process of refrigeration system under given heat load and the refrigeration power consumption of refrigeration system are predicted, and the refrigeration system prediction model is constructed, and the refrigeration system prediction model constructed is trained using the sample data set;The operation of refrigeration system is simulated using the trained refrigeration system prediction model, to determine the shutdown temperature setting value that makes the cumulative energy consumption lowest, to realize the optimization of shutdown temperature setting.The application can predict the start-stop process and power consumption of refrigeration system in a period of time in the future, and reduce the overall energy consumption of refrigeration system by optimizing the shutdown temperature configuration of refrigeration system.

Description

Technical Field

[0001] This invention relates to the fields of computer simulation and optimization, and green computing technology, and in particular to a method for predicting and optimizing shutdown temperature settings for data center cooling systems. Background Technology

[0002] With the development of information technology, the construction of data centers is in a period of rapid development. The operation of data centers requires a safe and reliable computer room environment. Servers generate a lot of heat during operation. In order to ensure the continuous and reliable operation of computer servers, air conditioning and cooling systems are usually installed. Maintaining stable temperature and humidity within strict tolerance range is crucial to the reliability of IT systems.

[0003] like Figure 1 As shown, the refrigeration system consists of indoor and outdoor parts. The main module of the indoor part is a liquid-air heat exchanger, which absorbs hot air generated by the production equipment and then transfers and dissipates the heat to the outside through pipes, condensers, and fans. The resulting cold air is then compressed and delivered indoors. As a common industrial facility, the refrigeration system is a typical cyclical, multi-stage system, operating as a partially observable 2-input, 3-output system. Intuitively, an increase in the heat load x1(t) (the total operating power of the server equipment) will cause an increase in the inlet temperature y3(t) of the refrigeration system. Furthermore, due to heat exchange between the refrigeration system and the outdoor environment, the ambient temperature x2(t) also affects the operation of the refrigeration system. Figure 2 As shown, in order to maintain the safe and stable operation of the production process, the refrigeration system needs to maintain the indoor intake air temperature within a pre-set upper and lower limit range. When the intake air temperature T... i Reaching the upper limit of the set value T high When the refrigeration system starts its compressor, the cooling capacity y2(t) produced by the refrigeration system increases, and the operating power y1(t) of the refrigeration system increases. When the inlet temperature T... i Below the shutdown setting value T low When the compressor is off, the refrigeration system enters standby mode, and y1(t) drops very low. Accompanying the refrigeration and standby states, the refrigeration system exhibits typical cyclical multi-stage characteristics. When the refrigeration system is in standby mode, the inlet temperature continuously rises, its cooling capacity is zero, and its cooling power is close to zero. When the refrigeration system is operating, the inlet temperature continuously decreases, and the cooling capacity and cooling power, after a rapid rise and oscillation, tend to stabilize, thus achieving stable refrigeration.

[0004] As competition intensifies in data centers and operating costs increase, statistics show that for every 100W increase in equipment power consumption, 10 to 20W is lost in power conversion, distribution, and transmission. Furthermore, the power consumption of air conditioning or cooling equipment can increase by as much as 120W, making the actual energy consumption of data centers far exceed the power consumption of the IT equipment itself.

[0005] For typical cooling systems, temperature sensors monitor ambient temperature to control the compressor's start and stop. When the temperature exceeds a certain upper threshold, the compressor starts and transfers heat from the room to the outside to cool the system. When the temperature drops to a lower threshold, the compressor shuts off, and the cooling system enters standby mode. The compressor is the most power-consuming device in the cooling system. Generally, appropriately increasing the lower temperature threshold can prevent overcooling and reduce energy consumption by decreasing compressor operating time. However, because starting the compressor generates 3-7 times the current of normal operation for a short period, an excessively high temperature threshold can cause the compressor to start and stop frequently, actually increasing the cooling system's power consumption. Currently, there is no technology in this field that can help data center operators accurately find the optimal shutdown temperature setpoint that achieves the best overall power consumption. Summary of the Invention

[0006] This invention provides a method for predicting and optimizing the shutdown temperature setting of a data center cooling system, so as to predict the start-up and shutdown process and power consumption of the cooling system over a period of time in the future, and reduce the overall energy consumption of the cooling system by optimizing the shutdown temperature configuration of the cooling system.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] On one hand, the present invention provides a method for optimizing the prediction and shutdown temperature settings of a data center cooling system, the method comprising:

[0009] Acquire time-series data of the refrigeration system during operation;

[0010] By combining a time series segmentation and clustering model with a manual annotation method, the time series data of the refrigeration system during operation are identified and classified in stages, and a sample dataset is constructed using the classified data.

[0011] A refrigeration system prediction model is constructed to predict the start-up and shutdown process and refrigeration power consumption of a refrigeration system under a given heat load, and the constructed refrigeration system prediction model is trained using the sample dataset; wherein, the input of the refrigeration system prediction model is the time series data of the refrigeration system during operation and the system input for a future period, and the output of the refrigeration system prediction model is the system output for a future period;

[0012] The operation of the refrigeration system is simulated using a trained predictive model. Based on the simulation results, the shutdown temperature setpoint of the refrigeration system that minimizes cumulative energy consumption is determined, thereby optimizing the shutdown temperature setting.

[0013] Furthermore, acquiring the time-series data during the operation of the refrigeration system includes:

[0014] Acquire operational data of the cooling system during operation; wherein, the operational data includes: server power, room temperature, server air inlet temperature, cooling power, and cooling capacity;

[0015] Missing values ​​in the running data are interpolated using cubic splines to form a time series with uniform sampling intervals.

[0016] Furthermore, the step of using a time-series segmentation clustering model combined with manual annotation to perform stage identification and classification on the acquired time-series data of the refrigeration system during operation includes:

[0017] Establish a semi-supervised multidimensional time series segmentation and clustering model;

[0018] On the original time series data, the operation process stage type of the refrigeration system to which some sequence intervals belong is manually labeled; wherein, the operation process stage type includes refrigeration shutdown, refrigeration startup first stage, refrigeration startup second stage, refrigeration stable cooling, and refrigeration shutdown process;

[0019] We use a constructed multidimensional time series segmentation and clustering model to learn manually labeled regions, and perform stage segmentation and category labeling on unlabeled regions of time series data based on the learning results.

[0020] Furthermore, the multidimensional time series segmentation clustering model clusters time series data based on the Topholtz inverse covariance matrix, and uses the Topholtz inverse covariance matrix to define a Markov random network for each cluster to describe the correlation of each dimension of the time series data during the operation of the refrigeration system over a period of time.

[0021] Furthermore, the step of using the constructed multidimensional time series segmentation and clustering model to learn manually labeled regions, and performing stage segmentation and category labeling on unlabeled regions of the time series data based on the learning results, includes:

[0022] The clustering results of each point in the time series data are obtained from manually labeled regions, and the Topholtz inverse covariance matrix in the time series segmentation clustering model is initialized using the alternating direction multiplier method.

[0023] The Expectation-Maximization (EM) algorithm is used to continuously update the parameters of the time series segmentation and clustering model and to achieve time series segmentation and clustering. In the EM algorithm, the M step is to update the Topholtz inverse covariance matrix using the alternating direction multiplier method given the clustering results of the time series data. In the E step, given the Topholtz inverse covariance matrix, a dynamic programming algorithm is used to find the minimum cost path, thereby segmenting and clustering the time series data.

[0024] Furthermore, the refrigeration system prediction model is an encoder-decoder framework structure; wherein,

[0025] The encoder takes as input the operating data of the refrigeration system over a period of time and outputs as the current state code of the system, which serves as the initial state required by the decoder. The initial state includes: predicted latent variables, the current stage of the system, and the duration of the current stage.

[0026] The decoder takes as input the state code given by the encoder and system input for a future period of time, and outputs as system output for a future period of time; wherein the system input includes server power and room temperature, and the system output includes inlet temperature, cooling power and cooling capacity.

[0027] Furthermore, both the encoder and the decoder are constructed from finite state machine ordinary differential equation networks.

[0028] Furthermore, the simulation of the refrigeration system operation using the trained refrigeration system prediction model, and the determination of the refrigeration system shutdown temperature setpoint that minimizes cumulative energy consumption based on the simulation results, includes:

[0029] The system cooling process under different shutdown temperature settings is simulated using a trained cooling system prediction model. The simulation shows the operation of the cooling system under different server heat loads, predicts the cumulative energy consumption of the cooling system over a period of time, and finds the shutdown temperature setting that minimizes the cumulative energy consumption of the system based on the simulation results.

[0030] In another aspect, the present invention also provides an electronic device comprising a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described method.

[0031] In another aspect, the present invention also provides a computer-readable storage medium storing at least one instruction that is loaded and executed by a processor to implement the above-described method.

[0032] The beneficial effects of the technical solution provided by this invention include at least the following:

[0033] This invention acquires a multidimensional time-series dataset of a refrigeration system; classifies and labels each stage of the dataset using a combination of manual methods and segmentation clustering models; establishes and trains a refrigeration system prediction model to predict and simulate the operation of the refrigeration system and calculate the power consumption over a future period; uses the refrigeration system prediction model to simulate the operation of the refrigeration system with different shutdown set temperature configurations, selects the optimal shutdown set temperature configuration, and thereby reduces the overall energy consumption of the refrigeration system by optimizing the system shutdown temperature configuration. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 It is a diagram showing the operation process and heat exchange of the refrigeration system;

[0036] Figure 2 This is a simplified diagram of the operating principle of a refrigeration system;

[0037] Figure 3 This is a schematic diagram of the execution flow of the data center cooling system prediction and shutdown temperature setting optimization method provided in the embodiments of the present invention;

[0038] Figure 4 This is a timing diagram of the output quantity of the refrigeration system dataset provided in an embodiment of the present invention;

[0039] Figure 5 This is a dynamic programming cost path diagram for segmenting and clustering time series data, provided in an embodiment of the present invention.

[0040] Figure 6 This is the encoder-decoder framework and corresponding data stream of the refrigeration system prediction model provided in this embodiment of the invention;

[0041] Figure 7 This is a schematic diagram of the Finite State Machine-Ordinary Differential Equation Network Model (DFA-ODENet) provided in an embodiment of the present invention;

[0042] Figure 8 This is a comparison chart of the actual output and model prediction output of the refrigeration system provided in an embodiment of the present invention;

[0043] Figure 9 This is a simulation diagram of the cooling power output of the refrigeration system provided in this embodiment of the invention under different shutdown temperature setpoints;

[0044] Figure 10 This is a schematic diagram illustrating the impact of the optimal shutdown temperature setting value on energy consumption under different thermal loads, as provided in the embodiments of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0046] First Embodiment

[0047] This embodiment provides a method for predicting and optimizing the shutdown temperature setting of a data center cooling system. This method can predict the start-up and shutdown process and power consumption of the cooling system over a future period, and reduce the overall energy consumption of the cooling system by optimizing the shutdown temperature configuration. This method can be implemented by an electronic device, such as a terminal or a server. The execution flow of this method is as follows: Figure 3 As shown, it includes the following steps:

[0048] S1, acquire time-series data of the refrigeration system during operation;

[0049] Specifically, in this embodiment, the implementation process of S1 is as follows:

[0050] S11: The operation of the refrigeration system is monitored by sensors to obtain operating data with periodic multi-stage characteristics during the operation of the refrigeration system, and the data is saved to generate a refrigeration system dataset.

[0051] The operational data includes: server power, room temperature, server air inlet temperature, cooling power, and cooling capacity; the cooling system's operational data can be obtained by calling the Seduce platform's API. The Seduce platform is an IoT platform for power and temperature management. This system collects sensor monitoring data at a frequency of 1 Hz, including: server power, room temperature, cooling system power, cooling capacity, and air inlet temperature data, such as... Figure 4 As shown, this is a time series of server air inlet temperature, cooling capacity, and cooling power extracted from the dataset, which clearly demonstrates that the server's operation process exhibits a cyclical, multi-stage characteristic.

[0052] The acquired operational data comes from different operating power periods of the server. Specifically, in this embodiment, the acquired operational data comes from periods when the server power (thermal load) is stably running at 1.7kW, 3.8kW, and 6.3kW. In this embodiment, the data is stored in different files according to different operating power, and saved as .csv files. The dataset length for thermal load 1.7kW is 8853s, the dataset length for thermal load 3.8kW is 9771s, and the dataset length for thermal load 6.3kW is 8418s.

[0053] S12 performs cubic spline interpolation on the missing values ​​in the dataset to form a time series with uniform sampling intervals.

[0054] It should be noted that the dataset contains a small number of missing values. Cubic spline interpolation is used, which causes minimal damage to the original sequence data, thus generating time series data with uniform intervals.

[0055] S2 uses a time series segmentation and clustering model combined with manual annotation to identify and classify the time series data of the refrigeration system during operation, and uses the classified data to construct a sample dataset.

[0056] Specifically, in this embodiment, the objective of S2 described above is to process the time-series data x of the refrigeration system. orig = [x1,x2,x3…x T Based on their dynamic characteristics, the system is divided into stages and categorized into k classes (k = 5 in this embodiment), where x orig Given an n-dimensional time series of duration T, to classify these T time points into k clusters, it's not enough to focus solely on the clustering of each individual point; the contextual information of that point is also crucial. This model uses a sliding window of size w to divide the original data into several small segments, defined as X = ..., X1, X2, X3...X... T For time point t, each segment X t For an nw-dimensional vector [x t-w+1 ,…,x t In this way, each point in the original sequence corresponds bidirectionally to a subsequence with a window size of w. The problem is transformed into clustering these subsequences X1…X T The problem to be solved is to process K inverse covariance matrices Θ = {Θ1, Θ2, ..., Θ...} k}, thus obtaining the result set P = {P1, P2, ..., P} k}, where P i The data is a set of multiple points, meaning that T points will be divided into k classes to obtain the classification results for each stage of the refrigeration system sequence data.

[0057] Based on the above, in this embodiment, the implementation process of S2 is as follows:

[0058] S21, Establish a semi-supervised multidimensional time series segmentation and clustering model;

[0059] The multidimensional time series segmentation clustering model clusters time series data based on the Topholz inverse covariance matrix. It uses the Topholz inverse covariance matrix to define the Markov random network of each cluster to describe the correlation of each dimension of the time series data during the operation of the refrigeration system over a period of time.

[0060] Using the Toploz inverse covariance matrix Θ i ∈R nw×nw Define each class:

[0061]

[0062] Among them, A (0) A (1),…,A (w-1) ∈R n×n Each A represents the relationship between the dimensions at the same point in time. The inverse covariance matrix reflects the independence between the variables, for example, A (1) ij This shows the correlation between dimension i at time t and dimension j at time t+1. A (2) The relationship between each dimension at time t and t+2 is shown. The inverse covariance matrix is ​​constrained to a Topols structure, thus defining a Markov random network for each cluster. Each Markov network corresponding to a cluster has w layers, which is the sliding window length defined above. The intra-layer represents the correlation of each dimension, and the inter-layer represents the correlation of each dimension at different time intervals. It is a time-invariant system, that is, the edge between layer 1 and layer 2 also exists between layer 2 and layer 3.

[0063] S22, On the original time series data, manually label the type of the operation process stage of the refrigeration system to which a certain sequence interval belongs; wherein, the type of operation process stage includes refrigeration shutdown, refrigeration startup first stage, refrigeration startup second stage, refrigeration stable cooling, and refrigeration shutdown process;

[0064] It should be noted that the analysis of the refrigeration system's dynamic characteristics revealed five stages: refrigeration unit shutdown, refrigeration unit startup stage 1, refrigeration unit startup stage 2, stable cooling, and refrigeration system standby. Each stage is characterized by the following: during refrigeration unit shutdown, the refrigeration system's operating power is 0W; during refrigeration unit startup stage 1, the operating power rapidly rises to its peak; during refrigeration unit startup stage 2, the operating power reaches its peak and then begins to decline; during stable cooling, the refrigerant system's cooling power exceeds 5kW, and the cooling capacity remains stable at 23300W; during refrigeration system standby, the cooling capacity drops to 0, and the refrigeration unit's power remains in the 200-300W range.

[0065] S23 utilizes the constructed multidimensional time series segmentation and clustering model to learn manually labeled regions, and performs stage segmentation and category labeling on unlabeled regions of time series data based on the learning results.

[0066] Specifically, the implementation process of S23 above is as follows:

[0067] S231: The clustering results of each point in the time series data are obtained from the manually labeled regions, and the ADMM algorithm (alternating direction multiplier method) is used to initialize the Topolaz inverse covariance matrix in the time series segmentation clustering model; the result set P is obtained from a manually labeled part of the refrigeration system dataset, and the Topolaz inverse covariance matrix Θ of the model is initialized through the M steps of the EM algorithm.

[0068] S232 utilizes the EM algorithm (Expectation-Maximization algorithm) to continuously update model parameters and achieve time series segmentation and clustering. The M-step of the EM algorithm updates the Topholtz inverse covariance matrix using the ADMM method, given the clustering results of the monitoring data time series. The E-step, given the Topholtz inverse covariance matrix, uses a dynamic programming algorithm to find the minimum cost path, thereby segmenting and clustering the monitoring data time series.

[0069] The E phase of the EM algorithm involves assigning points to each cluster and obtaining the result set P, as detailed below:

[0070] Use a dynamic programming algorithm to solve each X t Assigned to a cluster. For example... Figure 5 As shown, the loss for each node is a negative log-likelihood function:

[0071]

[0072] Where -ll(i,j) is the negative log-likelihood of assigning point i to class j, and the edge loss is β. Dynamic programming can be understood as having a total of T sequences, and assigning each sequence to different classes in sequence. If two adjacent sequences are assigned to different classes, a loss β will be generated; if they are in the same class, the loss will be 0. The goal of the whole process is to minimize the total cost and find the minimum cost path. This forms the classic pipeline scheduling problem, which is solved using the dynamic programming algorithm. Points are assigned to each cluster, and the result set P is obtained.

[0073] In the EM algorithm, step M involves updating the cluster parameter Θ given the result set P, as detailed below.

[0074] Solve the following optimization problem.

[0075]

[0076] in, For the indicator function, check whether adjacent points are assigned to the same cluster. Rewrite the log-likelihood part of the above formula as follows:

[0077]

[0078] Among them, |P i | is the number of cluster points assigned to S. i C is the empirical covariance matrix of these points, and C is a Θ-independent matrix. i The constant is given. The above problem is transformed into an optimization problem.

[0079]

[0080]

[0081] Where λ is a regularization parameter used to sparse the inverse covariance Θ i Adding a consensus variable Z transforms the problem into an ADMM problem.

[0082]

[0083] Where ρ is the penalty parameter, U is the dual variable, and Z = Θ in the above equation.

[0084] The ADMM algorithm consists of the following three steps, which are repeated until convergence is achieved, thus completing the update of Θ.

[0085]

[0086]

[0087] (c)U k+1 :=U k +(Θ k+1 -Z k+1 )

[0088] Where K is the current iteration round.

[0089] Repeat the above two EM steps until convergence is achieved, thus realizing multidimensional time series segmentation and clustering, and completing the identification and classification of each stage of the sequence monitoring data of the refrigeration system.

[0090] S3, construct a refrigeration system prediction model to predict the start-up and shutdown process and refrigeration power consumption of the refrigeration system under a given heat load, and train the constructed refrigeration system prediction model using the sample dataset; wherein, the input of the refrigeration system prediction model is the time series data of the refrigeration system during operation and the system input for a future period, and the output of the refrigeration system prediction model is the system output for a future period.

[0091] In this embodiment, the refrigeration system prediction model is an encoder-decoder framework, with both the encoder and decoder constructed from finite state machine ordinary differential equation networks. For the encoder module, given monitoring data of the refrigeration system over a past period (server power, room temperature, server inlet temperature, refrigeration power, and refrigeration capacity), the model can provide the current state code of the system as input to the decoder module. For the decoder module, in addition to the state code provided by the encoder, given system inputs over a future period (server power and room temperature), the model can predict the system outputs (inlet temperature, refrigeration power, and refrigeration capacity) during that period, thereby simulating the operating power consumption of the refrigeration system.

[0092] Based on the above, in this embodiment, the implementation process of S3 is as follows:

[0093] S31 introduces an encoder-decoder framework, such as Figure 6 As shown, the encoder and decoder are constructed using two finite state machine-ordinary differential equation network modules, DFA-ODENets, respectively.

[0094] The time series data input to the model includes conditional ranges. and prediction range First, all data within the specified range are preprocessed and converted into continuous time series. Then, an encoder is used to process the conditional range [t1, t2]. I The data within is encoded to obtain the initial state required by the DFA-ODENets decoder, i.e., t. I The state S(t) at time t I )=[h(t I ),s(t I ),T(t I )] T , where h(t) I ),s(t I ),T(t I The three terms are the predicted latent variables, the current stage of the system, and the duration of the current stage. Then, during the prediction phase, the DFA-ODENets-based decoder will determine the initial state S(t). I ) and system external input prediction system in the range [t I ,t I+L The output is as follows:

[0095] The framework takes continuous input data, and the encoder solves for the initial state during the encoding phase. Then, based on the solved initial state, the framework uses the decoder to solve an ordinary differential equation network, predicting the future system output from the input within the prediction range; specifically, within the conditional range [t1, t2]... I In the encoding stage, we combine the continuous-time signal cooling system output y(t) and the cooling system input x(t) as the input to the DFA-ODENet encoder to generate the initial state. As shown in the following formula:

[0096]

[0097] The initial state at time t1 is defined as: S(t1) = [h(t1) = 0, T(t1) = 0, s(t1)] T Based on the intermediate state values ​​obtained from the solution. You can obtain the status. Where h(t) I ) is from the distribution mean is The gradients were sampled from a diagonal multivariate Gaussian distribution with a covariance of I. The sampling process used reparameterization to construct a computational graph for gradient propagation and to reduce the variance of gradient estimates during training.

[0098] Sampling from the probability distribution is equivalent to treating h(t) as a latent variable in the model generated from the conditional sequence, and using the system output and input sequences to infer the approximate posterior distribution of the latent variable. Next, given the initial state S(t) of the decoding stage... I ) and prediction range [t I ,t I+L ]Continuous-time system input x([t I ,t I+L The predicted state within the prediction range is obtained by using the DFA-ODENet decoder.

[0099]

[0100] Then solve for any time t i System output

[0101] During the initial state determination process in the encoding phase, the encoder does not introduce a phase transition predictor, but instead achieves accurate phase transitions based on the actual system output. In the decoding phase, the duration and transition of the system phases are determined by the phase duration predictor.

[0102] like Figure 7 As shown, the finite state machine-ordinary differential equation network module of the refrigeration system prediction model mainly includes N ordinary differential equation network (ODENet) models. According to the state variables, only one ODENet module is activated at any given time for system prediction.

[0103] DFA-ODENet can also be seen as an extended version of ODENet, such as... Figure 7 As shown, compared to ordinary ODENet which only predicts the derivative of the hidden state, DFA-ODENet extends the hidden state to three parameters, thus covering changes in both the output variable and the stage variable in the system. The calculation and updating of the hidden state are in the continuous time domain, represented as S(t) = [h(t), s(t), T(t)], where the three terms are the predicted hidden variable, the current stage of the system, and the duration of the current stage, respectively. The solution process can be viewed as calculating the solution of the differential equation given a time point t and an infinitesimal time step dt, as shown in the following formula:

[0104]

[0105] in, The derivative of the latent variables is used by the duration predictor to estimate the duration of the current stage. Calculated by a stable gated cyclic unit (GRU) network:

[0106]

[0107] in, t is a learnable parameter in the model, and μ(t) is the time interval between t and t+1.

[0108] Solve for the hidden state h(t) of the system at any time, and then use it as the input of a fully connected neural network NN to predict the system output y(t) at the current time.

[0109]

[0110] The duration T(t) of the system in the current stage and the system input x(t) are used as feature inputs to assist in calculating the derivative of h(t) and predicting y(t). These are the learnable parameters in the model.

[0111] Furthermore, such as Figure 7 As shown, the refrigeration system prediction model also integrates a stage transition predictor. The stage transition predictor can identify the current stage of the refrigeration system and act as a switcher for multiple ODENets, assigning appropriate ODENets to predict the dynamic characteristics within each stage.

[0112] A learnable duration predictor is introduced into the stage transition predictor. The duration predictor estimates the duration of the current stage when predicting the system output, and helps the model determine whether to transition to the next stage. The duration predictor is constructed as a regular multilayer perceptron, using the current hidden states h(t) and x(t) as input:

[0113]

[0114] Where, φ s(t) For learnable parameters, when satisfying When the model transitions to the next stage (s(t)+1)mod N and resets T(t) to 0, in this embodiment N=5.

[0115] S32, the sample dataset is divided into a training set and a validation set. The training set is used to train the refrigeration system prediction model, and the validation set is used to evaluate the model and find the optimal model hyperparameters and training parameters.

[0116] In this embodiment, for the three datasets with different heat loads, the complete sequence length in each dataset is approximately 8000-10000, the sampling time points are continuous and non-uniform, the average sampling frequency is approximately 1 sequence / second, and the corresponding time length of the sequence is approximately 8000s-10000s. The first 50% of the data in each dataset is used for model training, and of the remaining data, 25% is used to build the validation set, and 25% is used for model testing. For the training and validation sets, this embodiment uses a sliding window of size 1600s to traverse the original sequences and generate training samples. For each training sample, the first 800s is the conditional range, used for the model to solve the initial state required for the prediction stage, and the last 800s is the prediction range, within which the model predicts the system output.

[0117] This embodiment defines the dataset time range as [t1:t]. I+L [[t1:t]] is divided into conditional ranges [t1:t] I ] and prediction range [t I :t I+L The model consists of two parts. The model is based on a given condition range [t1:t...]. I The system input and output under [t] generate the initial state, and then according to the prediction range [t] I :t I+L Given the future system input, the system output within the prediction range is predicted, as shown in the formula: in the formula This indicates the predicted output. This represents the system's historical output within the specified conditions. This represents the system input over the entire time span, including both historical and future inputs. ζ represents the model parameters.

[0118] When training a predictive model for a refrigeration system, the parameters trained in the encoder-decoder structure consist of three parts: ζ = {Θ} d ,Θ e Φ represents the N ODENets in the encoder, the N ODENets in the decoder, and the stage transition predictor in the decoder, respectively. To train the encoder and decoder parameters in an end-to-end manner, the model loss function to be optimized includes two parts: the error loss from predicting the system output. Error loss of the duration of each stage estimated by the duration predictor Where λ is the equilibrium and The weight parameters, for Because the model will h(t) ISince the evidence lower bound (ELBO) is used as a latent variable for posterior inference, the variational Bayes method can be adopted to maximize the likelihood of the system output by taking the maximization of the evidence lower bound (ELBO) of the system's observed output as the objective of model training.

[0119]

[0120] Define the hidden variable p(h(t) I The prior distribution of )) follows a normal distribution Normal(h(t) I ); 0, I), the posterior distribution of the model estimate It follows a diagonal multivariate Gaussian distribution. Therefore, the KL divergence term can be simplified to for h(t) I The regularization term is used for the decoder part. For the decoder part, we define the generative model. Given a normal distribution with a fixed covariance matrix, the mean of the distribution is defined as the expected value of the system output predicted by the DFA-ODENet decoder, maximizing the reconstruction likelihood. This can be simplified to minimizing the predicted system output value. Compared with the actual system output between L 2 Distance. In summary, the loss of the prediction system output and the loss of the latent variable posterior estimation are expressed as:

[0121]

[0122] in, and y(t) i () represent time t i The actual output and predicted output of the refrigeration system.

[0123] Furthermore, during the training process of the stage duration predictor, its error loss... Defined as the squared error between the predicted duration and the actual duration of each stage during training:

[0124]

[0125] in, Representative stage s(t) i The predicted duration of ) For s(t) i Actual duration. Calculation. We need to find t i The interval boundary t of the current stage l and t r . t l For [t] IWithin the range of ,∞]], the condition s(t) is satisfied. r )=s(t i And s(t) r +dt)≠s(t i The minimum value of ). T l For (-∞,t) i Within the range of ), condition s(t) is satisfied. l )=s(t i And s(t) l -dt)≠s(t i The maximum value of ) can be obtained.

[0126] S4 uses a trained refrigeration system prediction model to simulate the operation of the refrigeration system, and determines the system shutdown temperature setpoint that minimizes cumulative energy consumption based on the simulation results, thereby optimizing the shutdown temperature setting.

[0127] It should be noted that the inlet air temperature setpoint is one of the key settings affecting the operation of the refrigeration system. Typically, the system needs to be set to a threshold temperature of the air entering the inlet (start-up temperature T). high and shutdown temperature T low To ensure a safe operating environment, the initial inlet temperature setpoint for most cooling systems is typically fixed, with configuration values ​​lower than those defined in ASHRAE TC9.9 (Data Center Power Equipment Thermal Guide and Best Practices), failing to adequately consider actual cooling needs. Taking the cooling system studied in this embodiment as an example, its temperature thresholds are 12°C and 20°C. The centralized inlet temperature varies periodically between these two thresholds. To improve energy efficiency and ensure production process safety, this embodiment keeps the start-up temperature constant, adjusting only the shutdown temperature to reduce energy consumption. However, determining the optimal temperature threshold is challenging. If the threshold is set too low, excessive cooling will waste significant electrical energy. If the threshold is set too high, the compressor's cooling restart process will be more frequent, resulting in more peak power per unit time. Therefore, this embodiment employs the following steps to optimize the shutdown temperature setting.

[0128] S41, calculate the power consumption of the refrigeration system over a period of time based on the future operating power predicted by the trained refrigeration system prediction model.

[0129] It should be noted that during the testing phase, unlike the construction of the training set, the complete sequences in the test set were not segmented. The first 800 seconds of the sequence were fed into the encoder model to generate the initial state, and the decoder module, given the sequence input, predicted the remaining part of the system output in an open-loop manner. Figure 8As shown, this compares the actual output and predicted results of the cooling system. The time series trends of the predicted and actual outputs are quite consistent. Next, we will compare the predicted results with the actual outputs in the test set to evaluate the model accuracy. The long-term cumulative results of the model's predicted power consumption will be evaluated. For datasets with different power levels, we use the trained model to predict the future power consumption of the cooling system within the prediction range [t]. I :t I+L The chiller power under (total 120 minutes) is then defined. Then, an evaluation window T is defined, and [t]... I ,t I+L Divided according to size T into For each sub-interval, the absolute percentage error (APE(T)) of the power consumption prediction result is calculated. Then, the average of the evaluation results for all intervals is taken to obtain the mean absolute percentage error (MAPE(T)), as shown in the following formula:

[0130]

[0131]

[0132] The absolute percentage error (APE) is the relative error between the integral of instantaneous power and the actual power consumption over a period of time [t1, t2]. y1(t) and y1(t) represent the actual and predicted outputs of the cooling system at time t, respectively. When T equals 30 minutes, the MAPE (Maximum Power Equivalent) of both the predicted and actual power consumption values ​​are within 5%, indicating that the model has high accuracy in predicting the long-term power consumption of the cooling system.

[0133] S42 uses a trained refrigeration system prediction model to simulate the system's cooling process under different shutdown temperature setpoints. The model uses temperature judgment instead of time-based prediction to implement the transition from the refrigeration unit's stable cooling phase to its shutdown phase. Specifically, when the refrigeration system's inlet temperature drops to the given shutdown temperature setpoint, the system transitions from the stable cooling phase to the shutdown phase. Data sets with average server loads of 1.7k, 3.8k, and 6.3k are used. The shutdown temperature setpoint is gradually increased from 12℃ to 18℃ at 0.5℃ intervals. The simulation demonstrates the refrigeration system's operation under different loads and temperature setpoints, calculating the cumulative energy consumption of the refrigeration system over a period of time under different server operating power levels. The optimal shutdown temperature setpoint is then selected to reduce the refrigeration system's power consumption. Figure 9 This illustrates the impact of different shutdown temperature setpoints on the operation of the refrigeration system. From left to right, the simulation results are for 12℃, 16℃, and 18℃. As the temperature setpoint increases from 12℃ to 18℃, the number of refrigeration system starts increases significantly within the same time interval, leading to a continuous increase in system power consumption due to frequent compressor starts. Figure 10 When the shutdown temperature setpoint is increased from 12℃ to 18℃, the total energy consumption of the refrigeration system first decreases and then gradually increases. This is due to the higher T... low This narrows the gap between the upper and lower limits, causing the compressor to restart more frequently. Although the restart time is short, the compressor's power consumption is extremely high during this period. Therefore, starting from a certain temperature, the energy consumption due to compressor restarts accounts for the majority of the total cooling energy consumption, after which the total energy consumption increases. For three datasets with different heat loads, Figure 10 The optimal T that minimizes total energy consumption was selected. low Based on the simulation results, the energy savings achieved by using the optimal shutdown temperature setpoint are shown in the table below, thus achieving the goal of energy conservation.

[0134] Table 2 Optimal shutdown temperature setpoints and corresponding energy savings ratios

[0135] Average heat load (W) 1.7k 3.8k 6.3k Optimal shutdown temperature setpoint (degrees Celsius) 16.5 16 15 Energy saving percentage (%) 13.15 10.74 18.20

[0136] In summary, this embodiment provides a method for predicting and optimizing the shutdown temperature setting of a data center cooling system. This method utilizes a time series segmentation and clustering model to identify and classify the stages of multi-dimensional time series data of the cooling system. Simultaneously, using a finite state machine ordinary differential equation network as the core component, a cooling system prediction simulation model based on an encoder-decoder architecture is constructed. This model can simulate the system operation process and predict the energy consumption of the cooling system under different shutdown temperature setpoints. At the same time, it can inversely optimize the shutdown temperature setpoint of the cooling system to determine the optimal shutdown temperature configuration, thereby effectively reducing the overall energy consumption of the cooling system.

[0137] Second Embodiment

[0138] This embodiment provides an electronic device, which includes a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment.

[0139] The electronic device can vary considerably depending on its configuration or performance, and may include one or more processors (central processing units, CPUs) and one or more memories, wherein the memories store at least one instruction that is loaded by the processor and executed in accordance with the above method.

[0140] Third Embodiment

[0141] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc. The instruction stored therein can be loaded and executed by a processor in a terminal.

[0142] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0143] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0145] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0146] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A method for predicting and optimizing shutdown temperature settings in a data center cooling system, characterized in that, include: The system acquires operational data of the refrigeration system during operation; it performs cubic spline interpolation on missing values ​​in the operational data to construct time series data of the refrigeration system during operation with uniform sampling intervals; the operational data includes server power, room temperature, server inlet temperature, refrigeration power, and refrigeration capacity. By combining a time series segmentation and clustering model with a manual annotation method, the time series data of the refrigeration system during operation are identified and classified in stages, and a sample dataset is constructed using the classified data. A refrigeration system prediction model is constructed to predict the start-up and shutdown process and cooling power consumption of a refrigeration system under a given heat load. The model is trained using the sample dataset. The refrigeration system prediction model employs an encoder-decoder framework. The encoder's input is the refrigeration system's past operating data, and its output is the system's current state code, which serves as the initial state required by the decoder. The initial state includes the predicted latent variables, the system's current stage, and the duration of the current stage. The decoder's input is the state code from the encoder and the system input for a future period, and its output is the system output for that future period. The system input includes server power and room temperature, and the system output includes inlet temperature, cooling power, and cooling capacity. The operation of the refrigeration system is simulated using a trained predictive model of the refrigeration system. Based on the simulation results, the shutdown temperature setpoint of the refrigeration system that minimizes cumulative energy consumption is determined, thereby optimizing the shutdown temperature setpoint. The method of using a time series segmentation and clustering model combined with manual annotation to identify and classify the time series data of the refrigeration system during operation includes: A semi-supervised multidimensional time series segmentation and clustering model is established. The multidimensional time series segmentation and clustering model clusters time series data based on the Topolaz inverse covariance matrix. The Topolaz inverse covariance matrix is ​​used to define the Markov random network of each cluster to describe the correlation of each dimension of the time series data during the operation of the refrigeration system over a period of time. On the original time series data, the operation process stage type of the refrigeration system to which some sequence intervals belong is manually labeled; wherein, the operation process stage type includes refrigeration shutdown, refrigeration startup first stage, refrigeration startup second stage, refrigeration stable cooling, and refrigeration shutdown process; The clustering results of each point in the time series data are obtained from manually labeled regions, and the Topholtz inverse covariance matrix in the time series segmentation clustering model is initialized using the alternating direction multiplier method. The Expectation-Maximization (EM) algorithm is used to continuously update the parameters of the time series segmentation and clustering model and to achieve time series segmentation and clustering. In the EM algorithm, the M step is to update the Topholtz inverse covariance matrix using the alternating direction multiplier method given the clustering results of the time series data. In the E step, given the Topholtz inverse covariance matrix, a dynamic programming algorithm is used to find the minimum cost path, thereby segmenting and clustering the time series data.

2. The data center cooling system prediction and shutdown temperature setting optimization method as described in claim 1, characterized in that, Both the encoder and the decoder are constructed from finite state machine ordinary differential equation networks.

3. The data center cooling system prediction and shutdown temperature setting optimization method as described in claim 1, characterized in that, The process involves simulating the operation of the refrigeration system using a trained predictive model, and determining the refrigeration system shutdown temperature setpoint that minimizes cumulative energy consumption based on the simulation results. This includes: The system cooling process under different shutdown temperature settings is simulated using a trained cooling system prediction model. The simulation shows the operation of the cooling system under different server heat loads, predicts the cumulative energy consumption of the cooling system over a period of time, and finds the shutdown temperature setting that minimizes the cumulative energy consumption of the system based on the simulation results.

Images