Prediction model training method for temperature control and temperature control method and system
A neural network-based prediction model for data centers adjusts cooling demand based on load power and temperature data, addressing inefficiencies in existing methods by achieving precise temperature control and energy savings.
Patent Information
- Application Number
- CN202310651875.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-02
- Publication Date
- 2024-12-03
AI Technical Summary
Existing temperature control methods in data centers struggle to quickly and efficiently adjust cooling demand in response to significant changes in load power, leading to energy inefficiency and temperature imbalances, particularly in single-cabinet data centers.
A neural network-based prediction model is trained using time series data of load power, cabinet ambient temperature, cold channel temperature, and hot channel temperature to predict air conditioning cooling temperature, allowing for rapid adjustments in cooling operations based on real-time environmental conditions.
The model enables precise temperature control with reduced energy consumption, effectively matching cooling capacity to load demands, thereby optimizing energy efficiency and reducing the PUE index.
Smart Images

Figure CN119065411A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of air conditioning control technology, to the field of single-cabinet data center air conditioning energy-saving control technology, and specifically to a training method for a prediction model for temperature control and a temperature control method, system and storage medium for a single-cabinet data center based on the prediction model. Background Art
[0002] The internal thermodynamic characteristics of the data center are the bridge between the energy consumption of information equipment and the energy consumption of cooling equipment. In the thermodynamic research of the data center, most of them use the method of computational fluid dynamics (CFD) to model the entire data center and optimize the internal layout of the data center. The CFD method can correctly estimate the distribution of the temperature of the data center, but the computational load is too large, so it can be used for the design of the data center. However, the use of CFD for temperature estimation in the real-time optimization control of the existing data center cannot meet the real-time requirements. For the air-conditioning system, there is a large heat capacity and hysteresis in itself, and because the channel is closed, the instantaneous load change of the load power will reflect the impact of the temperature in the channel faster. The passive control method can easily cause the air-conditioning refrigeration to fail to keep up with the changes in the server load in time, and the cold and heat sources cannot be accurately matched, which makes the energy-saving effect less than ideal. Therefore, the traditional passive control average cooling method can hardly meet the temperature control requirements of the current micro-module computer room.
[0003] In the prior art, the application of artificial intelligence technology in data centers to optimize the setting of refrigeration system parameters has been studied. The method of optimizing the PUE index using a neural network model has been improved, the feature dimension has been increased, and the prediction accuracy has been improved. The machine learning method described above relies on the experience of on-site experts. It takes a long time for the model to be continuously optimized to obtain a better PUE operating state, and the implementation process and conclusion judgment are difficult. Another system and method for implementing energy-saving temperature control in data centers using artificial intelligence proposes combining deep neural networks and deep reinforcement learning to predict the temperature comfort under multiple influencing factors, and then find the optimal adjustment plan. Its deep neural network uses a two-layer hidden layer feedforward deep neural network to predict the appropriate temperature. The input includes air temperature and humidity, average temperature, wind speed, and change speed. The output is the predicted temperature comfort value, and the adjustment plan of the HVAC system is obtained through the model algorithm. This method does not consider the temperature distribution of the entire data center, but only considers the changes in indoor and outdoor air temperature and humidity in the computer room, ignoring the temperature impact caused by the heating of the equipment load, and the adjustment plan is relatively passive. Especially in the case of large changes in load power, the existing method cannot achieve overall balance in a short time, and its hysteresis easily leads to the occurrence of exceeding the temperature boundary conditions.
[0004] Therefore, a temperature control method is needed that can adjust the cooling demand quickly and with low power consumption when the load power changes significantly, so that the cooling capacity approaches the heat generated by the load. Summary of the invention
[0005] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a training method for a prediction model for temperature control and a temperature control method, system and storage medium for a single-cabinet data center based on the prediction model. By perceiving the environmental status of the data center and the changes in the equipment load power, the established prediction model is used to adjust the air conditioning refrigeration temperature to achieve optimal energy efficiency.
[0006] According to a first aspect, the present invention provides a training method for a prediction model for temperature control of a single-cabinet data center, comprising: obtaining time series data of load power, cabinet ambient temperature, cold channel temperature, hot channel temperature and air conditioning cooling temperature at historical moments at preset time intervals and preset time windows as training samples; wherein the time series data is generated during the historical temperature control process of the single-cabinet data center, the load power, the cabinet ambient temperature, the cold channel temperature, and the hot channel temperature are feature data used as inputs of a neural network model, and the air conditioning cooling temperature is a label used as an expected output value; iteratively optimizing the weights and bias parameters of the neural network model according to the data in the training samples until the error loss is lower than a preset error threshold or the number of iterations reaches a preset iteration threshold; and using the trained neural network model as a prediction model for temperature control of the single-cabinet data center.
[0007] Preferably, time series data of load power, cabinet ambient temperature, cold channel temperature, hot channel temperature and air conditioning cooling temperature at historical moments are obtained at preset time intervals and preset time windows as training samples, including: for the time series data of load power, cabinet ambient temperature, cold channel temperature, hot channel temperature and air conditioning cooling temperature at historical moments obtained at a first preset time interval and a first preset time window, resampling them at a second preset time interval and within a second preset time window to construct a single resampled time series data; all the resampled time series data are used as training samples; wherein the first preset time interval is smaller than the second preset time interval, and the first preset time window is smaller than the second preset time window.
[0008] Preferably, the historical temperature control process of the single-cabinet data center is a historical process of controlling the temperature of the single-cabinet data center by an adaptive control method.
[0009] Preferably, the neural network model is an error back propagation neural network.
[0010] According to a second aspect, the present invention provides a temperature control method for a single-cabinet data center, comprising: obtaining the load power, cabinet ambient temperature, cold channel temperature, and hot channel temperature of the single-cabinet data center at a current moment; using a prediction model obtained according to any method in the first aspect, outputting the air conditioning cooling temperature at the next moment according to the load power, cabinet ambient temperature, cold channel temperature, and hot channel temperature at the current moment; calculating the cooling demand based on the air conditioning cooling temperature at the next moment and performing a cooling operation.
[0011] According to the third aspect, the present invention provides a temperature control method for a single-cabinet data center, comprising: obtaining the load power of the single-cabinet data center at the current moment, and calculating the fluctuation rate of the load power; if the fluctuation rate exceeds a preset fluctuation rate threshold, executing the temperature control method of the single-cabinet data center based on a neural network in the second aspect; if the fluctuation rate does not exceed the preset fluctuation rate threshold, using an adaptive control method to control the temperature of the single-cabinet data center.
[0012] Preferably, in the process of executing the temperature control method of a single-cabinet data center based on a neural network, it also includes: if the cold channel temperature at the next moment exceeds the preset cold channel temperature range, or the hot channel temperature at the next moment exceeds the preset hot channel temperature range, then an adaptive control method is used to control the temperature of the single-cabinet data center.
[0013] Preferably, in the process of executing the temperature control method of a single-cabinet data center based on a neural network, it also includes: if the temperature control method of a single-cabinet data center based on a neural network has been continuously executed for a time length exceeding a preset time length, then an adaptive control method is used to control the temperature of the single-cabinet data center.
[0014] Preferably, the temperature of the single-cabinet data center is controlled by an adaptive control method, including: obtaining the cold channel temperature, hot channel temperature, and operating air conditioning cooling temperature of the single-cabinet data center at the current moment; calculating the air conditioning cooling temperature at the next moment according to a preset cold channel temperature range, a preset hot channel temperature range, the load power at the current moment, the cold channel temperature, the hot channel temperature, and the operating air conditioning cooling temperature; calculating the cooling demand based on the air conditioning cooling temperature at the next moment and performing a cooling operation.
[0015] Preferably, in the return air control mode, the air conditioning cooling temperature is the return air setting temperature; or in the supply air control mode, the air conditioning cooling temperature is the supply air setting temperature.
[0016] According to a fourth aspect, the present invention provides a temperature control system for a single-cabinet data center, comprising: a temperature acquisition module, a power acquisition module and a temperature control module; the temperature acquisition module is used to collect the cold channel temperature, the hot channel temperature, the running air-conditioning cooling temperature, and the cabinet ambient temperature; the power acquisition module is used to collect load power; the temperature control module is used to execute the temperature control method as described in any one of the second and third aspects.
[0017] According to a fifth aspect, the present invention provides a storage medium storing computer executable instructions, which, when loaded and executed by a processor, implement the steps of any method in the first to third aspects.
[0018] Compared with the prior art, the advantages of the present invention are: based on a large amount of measured data obtained when the data center air conditioner achieves the optimal energy efficiency operating state under the adaptive control method, a temperature prediction model based on a neural network is constructed that fully considers the temperature field characteristics of the data center, and the nonlinear relationship between the load power, hot channel temperature, cold channel temperature and cabinet ambient temperature and the air conditioning cooling temperature is obtained. Then, in the scenario where the load power undergoes a large wave, a temperature control method based on the prediction model can be used to perform constant temperature control using the predicted air conditioning cooling temperature. While quickly, accurately and effectively meeting the cabinet temperature requirements, the overall energy consumption of the air conditioner is reduced, thereby reducing the PUE index. The temperature can be intelligently controlled under conditions such as adaptive different geographical locations, different operating times, different rack deployments and loads, to achieve precise matching of cooling capacity, environment and load, and achieve energy-saving goals. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:
[0020] Figure 1 A deployment diagram of a control system in a cold channel of a micro-module data center according to an embodiment of the present invention;
[0021] Figure 2 A structural diagram of an air conditioning cooling temperature prediction model based on a neural network according to an embodiment of the present invention;
[0022] Figure 3 A schematic diagram of a hidden layer node model according to an embodiment of the present invention;
[0023] Figure 4 A schematic diagram of back propagation optimization of weights and bias parameters of a BP neural network according to an embodiment of the present invention;
[0024] Figure 5 A schematic diagram of a temperature control architecture based on a prediction model according to an embodiment of the present invention;
[0025] Figure 6 A schematic diagram of a process of temperature control based on a prediction model according to an embodiment of the present invention;
[0026] Figure 7 It is a schematic diagram of the adaptive control effect after the load power changes from 0 to 3kw in the air supply control mode;
[0027] Figure 8 It is a schematic diagram of the temperature control effect based on the prediction model after the load power changes from 0 to 3 kW in the air supply control mode according to an embodiment of the present invention. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail by specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0029] The micro-module data center is mainly composed of cabinets, enclosed passages, power supply and distribution systems, refrigeration systems, intelligent systems, integrated wiring and fire protection systems. Because it divides large data centers into several independent areas, it has the advantages of rapid deployment, flexible expansion, stability and reliability, green energy saving, intelligent management and efficient operation, etc. It meets the needs of high power density and low power consumption in the era of big data and cloud computing, and has become a hot spot for data center construction. In the pursuit of energy efficiency optimization of the refrigeration system, the optimal energy usage efficiency (PUE, Power Usage Effectiveness) is used as the energy-saving control target of air conditioning cooling. The optimal setting value of the air conditioning cooling temperature is determined under the constraint of meeting the temperature boundary conditions, and the air conditioning cooling is dynamically controlled. While accurately and effectively meeting the cabinet temperature requirements, it can produce a significant optimization effect on the PUE of the refrigeration system and even the entire data center.
[0030] Figure 1 FIG. 1 is a diagram of a deployment of a control system in a cold channel of a micromodule data center according to an embodiment of the present invention. Figure 1 As shown, three detection points are set in a single cabinet, among which the air inlet area temperature sensor 101 is set at the air inlet of the cabinet to collect the temperature of the cold channel or the temperature T of the air inlet area of the cabinet. inlet However, according to other embodiments of the present invention, it can also be set at other locations where the temperature of the cold air entering the cold channel or the cabinet can be collected. The air outlet area temperature sensor 102 is set at the air outlet of the cabinet to collect the temperature of the hot channel or the temperature T of the air outlet area of the cabinet. outletHowever, according to other embodiments of the present invention, it can also be set at other locations where the temperature of the air coming out of the cabinet can be collected to characterize the temperature of the air coming out of the cabinet at the first time. The ambient temperature sensor 103 is set at the top of the cold channel to collect the cabinet ambient temperature T env ; However, according to other embodiments of the present invention, the ambient temperature sensor 103 may also be arranged on the cabinet body or other locations where data for characterizing the ambient temperature of the cabinet can be collected. The sensors 101-103 send the collected temperature information to the monitoring management platform 104 via wired or wireless means, and the monitoring management platform 104 is responsible for monitoring and recording the operation data and executing a method for controlling the cooling of the air conditioner.
[0031] According to the "GB 50174-2017 Data Center Design Specification", the ambient temperature of the data center needs to meet the requirements of "the temperature of the cold channel or cabinet air inlet area is 18℃~27℃, and the ambient temperature and relative humidity of the main computer room (when shut down) are 5℃~45℃. In addition, the precision air supply setting range is 13℃~25℃, and the suitable temperature range of the UPS lead-acid battery pack is -15℃~40℃". The temperature boundary conditions for the safe operation of a single cabinet data center can be obtained by sorting out:
[0032] During operation, that is, the load power P dev >0, 18℃ <T inlet <27℃ and 5℃ <T outlet <40℃;
[0033] When the machine is stopped, the load power P dev =0, 5℃ <T inlet =T outlet <40℃;
[0034] The adjustment range of the return air temperature setting of the air conditioner is 16℃ <T return,set <40℃;
[0035] Air conditioning air supply set temperature adjustment range, 13℃ <T supply,set <25℃;
[0036] Among them, T inlet is the cold aisle temperature or the temperature of the cabinet air inlet area (hereinafter referred to as the cold aisle temperature), T outlet is the hot aisle temperature or the temperature of the cabinet outlet area (hereinafter referred to as hot aisle temperature), T return,set Set the return air temperature, T supply,set Set the temperature for the supply air. Load power P dev It mainly refers to the load power of IT equipment in the data center, and can also include the load power of other equipment in the data center except refrigeration equipment.
[0037] In order to better understand the temperature control method of the present invention, the temperature control method of adaptive adjustment in the return air control and supply air control modes (hereinafter referred to as the adaptive control method) is first introduced below.
[0038] 1. Return air control
[0039] Return air temperature control refers to the use of the temperature value collected by the temperature sensor on the return air side of the unit to participate in the control, compare the return air temperature of the air conditioner with the set return air set temperature, calculate the cooling demand to control the unit's capacity output and the on-demand actions of other components to adjust T inlet and T outlet , so that both meet the above temperature boundary conditions. The cooling demand is related to the return air temperature, return air set temperature, temperature dead zone, and temperature proportional band of the air conditioner, namely:
[0040] Cooling demand = f(return air temperature, return air set temperature, temperature dead zone, temperature proportional band) (1)
[0041] From the above formula, we can know that the cooling output of precision air conditioner can be adjusted by controlling the return air set temperature. Therefore, by dynamically adjusting the return air set temperature, the cooling capacity can be dynamically controlled. The principle of adaptive dynamic adjustment of return air set temperature algorithm is as follows:
[0042] T return,set (t+1)=T′ return,set (t)+k p (T cool_max -T inlet (t)) (2)
[0043] T return,set (t+1)=T′ return,set (t)+k p (T hot_max -T outlet (t)) (3)
[0044] Among them, T cool_max T is the temperature boundary condition of the data center. inlet The upper limit of the range, T hot_max The temperature boundary condition of the data center is T outlet The upper limit of the range, T outlet (t) is the hot channel temperature at time t, T inlet (t) is the cold channel temperature at time t, T′ return,set (t) is the return air set temperature at time t, T return,set (t+1) is the calculated return air set temperature at time t+1; k p It is the preset temperature proportional band.
[0045] Take the temperature boundary condition of the data center as follows: load power P dev >0, 18℃ <T inlet <27℃ and 5℃ <T outlet <40℃; load power P dev =0, 5℃ <T inlet =T outlet <40℃, the adjustment range of the return air temperature setting of the air conditioner is 16℃ <T return,set Taking <40℃ as an example, the adaptive control method under return air control is explained.
[0046] The sensors collect the single-cabinet data center operation data at the minute level, and then adjust the return air set temperature of the air conditioner every minute according to the cooling requirements under different operating conditions, as follows:
[0047] 1) Collect the working condition information at the current time t, including: load power P at time t dev (t), T outlet (t), T inlet (t) and T′ return,set (t).
[0048] The average temperature of multiple hot channels between time t-1 and time t is taken as T outlet (t); take the average value of multiple cold channel temperatures between time t-1 and time t as T inlet (t). In addition, since T′ return,set The value of (t) may be calculated or directly set, so when calculating the return air set temperature at time t+1, it is necessary to read the return air set temperature running at time t from the database or from the communication module of the air conditioner, instead of directly using the return air set temperature calculated at time t.
[0049] 2) In P dev (t)>0, that is, in the server running scenario:
[0050] If T inlet (t)<27℃, then calculate the return air set temperature T at time t+1 according to the following method return,set (t+1):
[0051] T return,set (t+1)=T′ return,set (t)+k p (40℃-T outlet (t))
[0052] According to the above temperature boundary conditions, T hot_max =40℃ as the high temperature boundary of the hot channel, dynamically adjust the return air set temperature to control the cooling demand to meet T outlet<40℃ environmental requirements;
[0053] Otherwise, the return air set temperature at time t+1 is T return,set (t+1) is calculated as:
[0054] T return,set (t+1)=T′ return,set (t)+k p (27℃-T inlet (t))
[0055] According to the above temperature boundary conditions, T cool_max = 27℃ as the cooling optimization target of cold channel temperature, and dynamically adjust the cold channel temperature by controlling the return air set temperature to meet T inlet Environmental requirement: <27℃.
[0056] In P dev (t) = 0, that is, in the server shutdown scenario, the return air set temperature T at time t+1 return,set (t+1) is calculated as:
[0057] T return,set (t+1)=T′ return,set (t)+k p (40c-T outlet (t))
[0058] According to the above temperature boundary conditions, T hot_max =40°C is used as the high temperature boundary of a single cabinet data center, and the return air set temperature is dynamically adjusted to meet T outlet <40℃ environment requirements, try to increase the return air set temperature to reduce cooling energy consumption. And when the load is shut down, the air conditioner has no cooling output.
[0059] 3) After calculating the return air set temperature T at time t+1 return,set After (t+1), it is also necessary to determine whether the calculated value is within the adjustment range of the return air set temperature of 16°C. <T return,set <40℃:
[0060] If T return,set (t+1)>40℃, then let T return,set (t+1) = 40°C;
[0061] If T return,set (t+1)<16℃, then let T return,set (t+1) = 16°C;
[0062] That is, the calculation result exceeds the return air setting temperature range, and the maximum or minimum return air setting value is output as the result.
[0063] 4) Based on the return air temperature of the air conditioner at time t and the return air setting temperature T at time t+1 obtained by the above calculation return,set (t+1), temperature dead zone, temperature proportional band k p According to formula 1), the cooling demand at time t+1 is calculated, and then the cooling operation is performed according to the calculated cooling demand to adjust T inlet and T outlet Make both meet the requirements of the temperature boundary conditions of the data center. Among them, the return air temperature of the air conditioner is collected by the air conditioner's own sensor, indicating the mixed air temperature of the return air channel (i.e., the hot channel). The temperature dead zone represents the error between the real-time value of the allowable return air temperature and the return air set temperature, that is, the error between the two is within the temperature dead zone, which can be regarded as having reached a regulation balance; the temperature dead zone can usually be set to -0.5°C ~ 0.5°C. The temperature dead zone and the temperature proportional band can be fixed values or values set at time t. The above calculation of the cooling demand based on the return air set temperature using formula (1) is a well-known technology in the field, and the present invention will not be repeated here.
[0064] 5) Return to step 1) and continue to dynamically adjust the return air set temperature at the next moment.
[0065] 2. Air supply control
[0066] Supply air temperature control refers to the use of the temperature value collected by the temperature sensor on the supply air side of the unit to participate in the control, compare the real-time value of the supply air side temperature value with the set supply air set temperature, and control the unit's capacity output and the on-demand actions of other components through the calculated cooling demand to adjust T inlet and T outlet , so that both meet the above temperature boundary conditions. The cooling demand is related to the air supply temperature, air supply set temperature, temperature dead zone, and temperature proportional band of the air conditioner, namely:
[0067] Cooling demand = f(supply air temperature, supply air set temperature, temperature dead zone, temperature proportional band) (4)
[0068] From the above formula, we can know that the cooling output of precision air conditioner can be adjusted by controlling the supply air set temperature. Therefore, the dynamic adjustment of the supply air set temperature can also dynamically control the cooling capacity. The principle of the algorithm for adaptive dynamic adjustment of the supply air set temperature is as follows:
[0069] T supply,set (t+1)=T′ supply,set (t)+k p (T cool_max -T inlet (t)) (5)
[0070] T supply,set (t+1)=T′ supply,set(t)+k p (T hot_max -T outlet (t)) (6)
[0071] Among them, T′ supply,set (t) is the supply air set temperature at time t, T supply,set (t+1) is the calculated supply air setting temperature at time t+1.
[0072] Still taking the temperature boundary condition of the data center as an example: load power P dev >0, 18℃ <T inlet <27℃ and 5℃ <T outlet <40℃, load power P dev =0, 5℃ <T inlet =T outlet <40℃, the adjustment range of air supply setting temperature is 13℃ <T supply,set Taking <25℃ as an example, the adaptive control method under air supply control is explained.
[0073] The sensors collect the single-cabinet data center operation data at the minute level, and then adjust the air supply set temperature of the air conditioner every minute according to the cooling requirements under different operating conditions. The details are as follows:
[0074] 6) Collect the working condition information at the current time t, including: load power P at time t dev (t), T outlet (t), T inlet (t) and T′ supply,set (t).
[0075] The average temperature of multiple hot channels between time t-1 and time t is taken as T outlet (t); take the average value of multiple cold channel temperatures between time t-1 and time t as T inlet (t). In addition, since T′ supply,set (t) may be calculated or directly set, so when calculating the supply air set temperature at time t+1, it is necessary to read the supply air set temperature running at time t from the database or from the communication module of the air conditioner, instead of directly using the supply air set temperature calculated at time t.
[0076] 7) In P dev (t)>0, that is, in the server running scenario:
[0077] If T outlet (t)<40℃, the air supply set temperature T at time t+1 is calculated as follows supply,set (t+1):
[0078] Tsupply,set (t+1)=T′ supply,set (t)+k p (27℃-T inlet (t))
[0079] According to the above temperature boundary conditions, T cool_max = 27℃ as the cooling optimization target of cold channel temperature, and dynamically adjust the cold channel temperature by controlling the supply air set temperature to meet T inlet Environmental requirements of <27℃;
[0080] Otherwise, the air supply setting temperature at time t+1 is T supply,set (t+1) is calculated as:
[0081] T supply,set (t+1)=T′ supply,set (t)+k p (40℃-T outlet (t))
[0082] According to the above temperature boundary conditions, T hot_max =40℃ as the high temperature boundary of the hot channel, dynamically adjust the supply air setting temperature to meet the cooling demand to meet T outlet <40℃ environmental requirement.
[0083] In P dev (t) = 0, that is, in the server downtime scenario:
[0084] If T outlet (t)>40℃, the air supply set temperature T at time t+1 is calculated as follows supply,set (t+1):
[0085] T supply,set (t+1)=T′ supply,set (t)+k p (40℃-T outlet (t))
[0086] According to the above temperature boundary conditions, T hot_max =40℃ as the high temperature boundary of the hot channel, and dynamically adjust the supply air setting temperature to meet T outlet <40℃ environment requirements, try to increase the air supply set temperature to reduce cooling energy consumption. And when the load is shut down, the air conditioner has no cooling output.
[0087] 8) Calculate the air supply set temperature T at time t+1 supply,set After (t+1), it is also necessary to determine whether the calculated value is within the adjustment range of the air supply set temperature of 13°C. <T supply,set <25℃:
[0088] If T supply,set (t+1)>25℃, then let T supply,set (t+1) = 25°C;
[0089] If T supply,set (t+1)<13℃, then let T supply,set (t+1) = 13°C;
[0090] That is, the calculation result exceeds the set air supply temperature range and is output with the maximum or minimum air supply set value as the result.
[0091] 9) Based on the air supply temperature of the air conditioner at time t, the calculated air supply setting temperature T at time t+1 supply,set (t+1), temperature dead zone, temperature proportional band k p According to formula 2), the cooling demand at time t+1 is calculated, and then the cooling operation is performed according to the calculated cooling demand to adjust T inlet and T outlet Make both meet the requirements of the temperature boundary conditions of the data center. Among them, the supply air temperature of the air conditioner is collected by the air conditioner's own sensor, indicating the cooling temperature of the supply air channel (i.e., the cold channel). The temperature dead zone represents the error between the real-time value of the allowable supply air temperature and the supply air set temperature, that is, the error between the two is within the temperature dead zone, which can be regarded as having reached a regulation balance; the temperature dead zone can usually be set to -0.5°C ~ 0.5°C. The temperature dead zone and the temperature proportional band can be fixed values or values set at time t. The above calculation of the cooling demand based on the supply air set temperature using formula (4) is a well-known technology in the field, and the present invention will not be repeated here.
[0092] 10) Return to step 6) and continue to dynamically adjust the air supply set temperature at the next moment.
[0093] From the above, it can be seen that the temperature control method based on the adaptive control method can make the air conditioner execute the adaptive control target at the minute level. Under the restriction of meeting the temperature boundary conditions of the data center, the air conditioner cooling temperature can be adjusted autonomously according to different operating environments, so that the cooling capacity and air volume output are adjusted according to the actual power of the load, and the high set temperature and less air conditioning cooling are always maintained to meet the load operating environment, greatly improve the matching degree with the actual load, and reduce the energy use efficiency (PUE) of the computer room, so as to achieve the purpose of reducing energy consumption. When the load power changes slightly, the adaptive control method can maintain the overall temperature balance in the cabinet, that is, the load heat dissipation level matches the air conditioner cooling demand. At this time, the cold channel temperature and the hot channel temperature change slightly, and the air conditioner set temperature is adjusted within a small range. However, when the load power changes greatly, the adaptive control method cannot achieve overall balance in a short time, and its hysteresis easily leads to the situation of exceeding the temperature boundary conditions. Therefore, a temperature control method is needed that can quickly and low-power adjust the cooling demand when the load power changes greatly, so that the cooling capacity approaches the heat generated by the load.
[0094] The inventors found in further research that when the air conditioner runs stably for a long enough time, the temperature field in the computer room reaches a cold and hot balance, the cooling capacity generated by the air conditioner approaches the heat generated by the rack load, and the hot channel temperature T outlet Equal to the air conditioning cooling temperature D and load power P dev , cold channel temperature T inlet , Cabinet ambient temperature T env The sum of the influence of factors such as , is expressed as:
[0095] T outlet =γ1T env +γ2P dev +γ3D+γ4T inlet +b T (7)
[0096] In formula (7), γ1 is the first influence coefficient of cabinet ambient temperature, γ2 is the first influence coefficient of load power, γ3 is the first influence coefficient of air conditioning cooling temperature, and γ4 is the first influence coefficient of cold channel temperature; b T It is the first offset term, covering other influencing factors. It should be noted that the air conditioning cooling temperature D under return air control represents the return air set temperature T return,set , which indicates the air supply set temperature T under air supply control supply,set , in the following content, the return air set temperature T under return air control return,set and the air supply set temperature T under air supply control supply,setThe combined value is recorded as the air conditioning cooling temperature D. The first influence coefficient and the first bias term are affected by factors such as the layout location of the data center, airflow characteristics, and load distribution. Generally speaking, the higher the load power, the higher the heat dissipation and temperature, the stronger the air conditioning cooling will be, and the faster the rack will cool down.
[0097] Furthermore, by transforming formula (7), we can obtain the demand relationship between the air conditioning cooling temperature and various environmental factors of the temperature field:
[0098] D=α1T env +α2P dev +α3T outlet +α4T inlet +b D (8)
[0099] In formula (8), α1 is the second influence coefficient of cabinet ambient temperature, α2 is the second influence coefficient of load power, α3 is the second influence coefficient of hot channel temperature, and α4 is the second influence coefficient of cold channel temperature; b D is the second bias parameter to cover other influencing factors.
[0100] Inside a single-cabinet data center, the cooling airflow flows from the air conditioner through the load area and then flows back to the vents, which is similar to the calculation process from input value to output value in a neural network; various environmental factors have influences on each other, and the strengths vary, similar to the connections between nodes in a neural network. Therefore, the inventors thought of using a neural network to model the temperature field of a data center, and training the model using a variety of operating data collected in the adaptive control method to obtain a nonlinear relationship between different influencing factors and the air conditioning cooling temperature, which is used to predict the air conditioning cooling temperature under different influencing factors.
[0101] The following combination Figure 2 The structure of an air conditioning cooling temperature prediction model based on a neural network proposed according to an embodiment of the present invention is described.
[0102] like Figure 2 As shown, according to one embodiment of the present invention, a neural network is used to construct an air conditioning cooling temperature prediction model based on formula (8). The neural network can be a generalized regression neural network, a probabilistic neural network, a radial basis function neural network, or other neural networks that can perform nonlinear function fitting. One embodiment of the present invention preferably uses an error back propagation (BP) neural network to reduce the loss through forward propagation calculation and back propagation to make the model converge. The prediction model is a three-layer fully connected neural network model including an input layer, a hidden layer, and an output layer. Among them, the input layer includes 4 nodes for characterizing 4 characteristic data load power P dev , Cabinet ambient temperature Tenv , cold channel temperature T inlet , hot channel temperature T outlet The nodes of the hidden layer are temperature field sensor (neuron) models, and their simplified structure shows that Figure 3 , the weights and bias parameters of the feature data can be adjusted through supervised training; according to one embodiment of the present invention, the number of hidden layer nodes can be determined according to the following formula:
[0103]
[0104] Wherein, p is the number of nodes in the hidden layer, m is the number of nodes in the input layer, n is the number of nodes in the output layer, and a is an adjustment constant, and the value range of a is 1 to 10; preferably, the number of nodes in the hidden layer is 6. The number of nodes in the output layer is 1, which is used to characterize the output air conditioning cooling temperature D. The nonlinear function relationship is fitted from the input layer to the hidden layer and from the hidden layer to the output layer through the activation function; in some embodiments, the activation function is preferably a sigmoid activation function, and its expression is:
[0105]
[0106] Among them, k represents the result calculated based on the input, weight and bias parameters of the previous layer. The weight from the input layer to the hidden layer is ω, the bias parameter is b1, and the weight from the hidden layer to the output layer is v, and the bias parameter is b2.
[0107] According to one embodiment of the present invention, the calculation process of the prediction model constructed by combining the above model structure with equations (8) and (10) from the input layer to the output layer is as follows:
[0108] 1) From input layer to hidden layer:
[0109] net1=ω T x+b1,h=g1(net1)
[0110] 2) From hidden layer to output layer:
[0111] net2=v T h+b2,y=g2(net2)
[0112] Where: x represents the input feature data, x1 is the cabinet ambient temperature, x2 is the load power, x3 is the hot aisle temperature, and x4 is the cold aisle temperature; ω ij represents the weight from the i-th node in the input layer to the j-th node in the hidden layer, i=1,2,...,m, j=1,2,...,n; in this embodiment, n=4, m=6; b 1jrepresents the bias parameter of the jth node from the input layer to the hidden layer; net1 is the network from the input layer to the hidden layer, net 1j represents the calculation result of the input, weight and bias parameters of the jth node from the input layer to the hidden layer; h represents the activation function from the input layer to the hidden layer, h j Represents the output of the jth node from the input layer to the hidden layer; v j represents the weight from the jth node of the hidden layer to the output layer; g1 is the activation function from the input layer to the hidden layer; g2 is the activation function from the hidden layer to the output layer; net2 is the network from the hidden layer to the output layer; y is the output of the output layer, which represents the cooling temperature of the air conditioner.
[0113] In summary, the calculation model from prediction model input to output is:
[0114] y=g2(net2)=g2(v T g1(ω T x+b1)+b2) (11)
[0115] like Figure 4 As shown, in the training phase, the BP neural network model needs to compare the error between the output of the output layer and the expected value, and adjust the weights and bias parameters of each model item through the back propagation process to minimize the error signal. According to one embodiment of the present invention, the optimization process of the model weights and bias parameters is:
[0116] 1) Initialize each weight and bias parameter. In some embodiments, a random number between 0 and 1 is used. The initial values of the weight ω from the input layer to the hidden layer, the bias parameter b1, the weight v from the hidden layer to the output layer, and the bias parameter b2 are denoted as ω, (0) 、b1 (0) 、v (0) 、b2 (0) .
[0117] 2) Activate forward propagation to obtain the output y of each layer, and calculate the error loss based on the loss function and expected value:
[0118]
[0119] Among them, D d is the true value of the air conditioning cooling temperature in the training sample.
[0120] 3) Calculate the error term of the output layer and the error term of the hidden layer. The gradient value or derivative of the weight and bias parameters can be expressed as dW i :
[0121]
[0122] Among them, W1=ω, W2=v, W3=b1, W4=b2.
[0123] ① Calculate the error term of the output layer, that is, calculate the gradient value or partial derivative of the loss function with respect to the output layer. According to the chain rule:
[0124]
[0125]
[0126] ② Calculate the error term of the hidden layer, that is, calculate the gradient value or partial derivative of the loss function with respect to the hidden layer. According to the chain rule:
[0127]
[0128]
[0129] 4) Update the weights and bias parameters in the model
[0130] The weight and bias parameters are updated using the Adam optimization algorithm. The update formula combined with equations (13) to (17) is as follows:
[0131]
[0132] V i (k) =β1V i (k-1) +(1-β1)dW i (k) (19)
[0133]
[0134] Where η represents the learning rate. By adjusting η, the pace of update can be adjusted. The preferred setting of η is 0.05; k represents the number of updates or iterations; W i (k) is W after the kth update i The value of V i (k) V after the kth update i The value of V i (0) =0, V i Represents the first-order momentum of the gradient of weight and bias parameters, V1 represents the first-order momentum of the gradient of ω, V2 represents the first-order momentum of the gradient of v, V3 represents the first-order momentum of the gradient of b1, and V4 represents the first-order momentum of the gradient of b2; S i (k) is S after the kth update i The value of Si (0) =0, S i Represents the second-order momentum of the gradient of weight and bias parameters, S1 represents the second-order momentum of the gradient of ω, S2 represents the second-order momentum of the gradient of v, S3 represents the second-order momentum of the gradient of b1, and S4 represents the second-order momentum of the gradient of b2; β1, β2, and ε are preset hyperparameters; preferably, β1=0.9, β2=0.999, and ε=10e-8.
[0135] The Adam optimization algorithm absorbs the advantages of the adaptive learning rate gradient descent algorithm and the momentum gradient descent algorithm, which can adapt to sparse gradients and alleviate the problem of gradient oscillation. The embodiment of the present invention adopts the Adam optimization algorithm to update weights and bias parameters. On the basis of the gradient descent method, by calculating the first-order momentum and second-order momentum of the gradient, independent adaptive learning rates are designed for different parameters to accelerate the convergence rate.
[0136] The weight and bias parameters of the model established by formula (11) are iteratively updated until the error loss E is less than the preset error threshold or the number of iterations reaches the preset iteration threshold, the model is determined to be converged, and the iteration is stopped. Preferably, the preset iteration threshold is 1000.
[0137] According to an embodiment of the present invention, based on the above model structure and training method, a method for constructing an air conditioning cooling temperature prediction model based on a neural network is proposed. Figure 5 As shown, it includes: collecting the load power P at the historical moment at a first preset time interval and a first preset time window dev , Cabinet ambient temperature T env , cold channel temperature T inlet , hot channel temperature T outletand the time series data of the air conditioning cooling temperature D as historical data, which is generated during the historical temperature control process of the single-cabinet data center. In some embodiments, the historical temperature control process of the single-cabinet data center includes: the historical process of controlling the temperature of the single-cabinet data center by an adaptive control method. The historical data covers high-efficiency optimized energy efficiency data under different load powers. Preferably, the data obtained during the stable operation of the air conditioner is high-efficiency and low-energy consumption data. During the stable operation of the air conditioner, that is, the overall temperature in the cabinet is balanced, the load heat dissipation level matches the air conditioning cooling demand, the air conditioning cooling temperature is adjusted within a smaller range, and the operating interval with small temperature changes in the cold channel and the hot channel. The first preset time interval is at the minute level, preferably 1 minute; the first preset time window is preferably 1 minute. In some embodiments, the historical data can also be resampled, and the historical data in the second preset time window is resampled at a second preset time interval to construct a single training sample. Among them, the resampling process includes but is not limited to: averaging, taking the median, and taking the mode; preferably, a single training sample is constructed by using the resampling process of averaging; the second preset time interval is preferably 5 minutes, and the second preset time window is preferably 5 minutes. Load power P dev , Cabinet ambient temperature T env , cold channel temperature T inlet , hot channel temperature T outlet The feature data is used as the input of the model, and the air conditioning cooling temperature D label is used as the expected output value (i.e., expected value); 70% of the training samples are used as the training set, and 30% are used as the test set. According to formula (11), Figure 2 The network model shown in the figure constructs a basic temperature field model, takes the feature data in the training set as input and the label as the expected value, and performs supervised training on the basic temperature field model according to the optimization method of the model weight and bias parameters until the loss is lower than the preset error threshold or the number of iterations reaches the preset iteration threshold, and the load power P is obtained. dev , Cabinet ambient temperature T env , cold channel temperature T inlet , hot channel temperature T outlet The nonlinear relationship between the air conditioning cooling temperature D. The performance of the trained optimal model is evaluated using the feature data in the test set as input and the label as the expected value. When the requirements are met, the air conditioning cooling temperature prediction model is obtained.
[0138] According to one embodiment of the present invention, Figure 5 A temperature control method based on a prediction model is proposed, such as Figure 6 , including: obtaining the current load power P of the single cabinet data center dev , Cabinet ambient temperature T env , cold channel temperature T inlet, hot channel temperature T outlet , using the above prediction model, according to the load power, cabinet ambient temperature, cold channel temperature, and hot channel temperature at the current moment, the air conditioning cooling temperature D at the next moment is output, that is, the return air set temperature (under return air control) or the supply air set temperature (under supply air control); based on the air conditioning cooling temperature D at the next moment, the cooling demand is calculated according to formula (1) or formula (4) and the cooling operation is performed. Compared with the adaptive control method, when the load power fluctuation exceeds the preset value, the temperature control method of the embodiment of the present invention can adjust the air conditioning cooling temperature based on historical experience to quickly adapt to the change of the current load power, find a reasonable air conditioning cooling temperature, and make the cooling capacity close to the heat generated by the load. Since the cold and hot balance is completed faster than the adaptive control method, the energy consumption will also be lower, achieving energy saving effect.
[0139] In some embodiments, when the load power fluctuation rate exceeds a preset fluctuation rate threshold (preferably, the preset fluctuation rate threshold is a fluctuation rate of >20% within 1 minute), a temperature control method based on a prediction model is executed within a preset time period for temperature control (preferably, the preset time period can be between 30 minutes and 40 minutes), and during the execution process, it is continuously detected whether the temperature boundary condition is exceeded. If the temperature boundary condition is exceeded or the temperature control based on the prediction model has exceeded the preset time period, the adaptive control method is switched to perform temperature control (i.e., according to the preset cold channel temperature range, the preset hot channel temperature range, and the collected load power, cold channel temperature, hot channel temperature, and the running air conditioning cooling temperature at the current moment, the air conditioning cooling temperature at the next moment is calculated), otherwise, the temperature control based on the prediction model is continued. It should be noted that the hot channel temperature at the current moment is the average value of the hot channel temperature in the time period from the previous moment to the current moment; the cold channel temperature at the current moment is the average value of the cold channel temperature in the time period from the previous moment to the current moment; and the return air setting temperature at the current moment is read from the database or the air conditioning communication module. The limitation of execution time and the continuous detection of whether the temperature boundary conditions are exceeded during the execution process are because if a working condition that has not appeared in the historical data occurs, the predicted air conditioning cooling temperature at the next moment may exceed the temperature boundary conditions. Limitation, detection and timely correction are carried out to prevent the problem of control failure after exceeding the temperature boundary conditions. The embodiment of the present invention can collect, manage, train, and infer a large amount of historical data after the air conditioning adaptive control achieves energy efficiency optimization under the current environment, load, cooling demand and other conditions, and obtain a temperature setting prediction model. Combined with the air conditioning adaptive adjustment and temperature field modeling, continuous optimization is carried out to obtain the control parameters with the lowest PUE, and collaborative control is used to achieve the optimal energy efficiency of the refrigeration system.
[0140] The beneficial effects of the air conditioning energy-saving optimization control method based on the adaptive control method and the temperature control method based on the prediction model of the present invention are respectively illustrated below with examples and diagrams.
[0141] In order to verify the boundary control capability of the air conditioning adaptive supply air temperature, when the load standby state is changed to the full load 3Kw, the air conditioning adaptive supply air temperature energy-saving algorithm based on micromodule control is used to autonomously adjust the air conditioning setting parameters of the cold channel and hot channel of the cabinet, so that the cooling capacity output is distributed on demand and the matching degree with the actual load is improved. Figure 7 The figure shows the adaptive energy-saving control effect when the load of a single-cabinet data center changes from standby to full load of 3Kw under the working condition of the outdoor ambient temperature of 33℃ on July 6, 2022. It can be concluded from the experimental figure that the micromodule control system calculates the supply air temperature setting value according to the adaptive control method to dynamically adjust the temperature of the cold channel and the hot channel to meet the temperature boundary conditions for safe operation, and the supply air temperature setting value finally makes the cabinet tend to cold and hot balance at 21.8℃. In the first 1 to 30 minutes, the adaptive control method has a lag in temperature regulation and exceeds the temperature constraint boundary, but finally reaches cold and hot balance after 80 minutes, the cold channel temperature and the hot channel temperature tend to be stable, and the supply air temperature setting value tends to be stable.
[0142] After the adaptive control method has been running for a long enough period of time, the parameters of the single-cabinet data center refrigeration system are optimized under different working conditions, and a temperature field model can be constructed to understand the influence of factors such as air conditioning refrigeration, data center layout, airflow pattern, and load rate changes on the inlet and outlet temperatures of the server rack. Using the historical time series data of air conditioning adaptive optimization as prior knowledge, neural networks are used for learning and training to establish a coupling relationship between the cooling demand characteristics and other related characteristics, and then the cooling demand required to maintain different conditions is calculated. The minute-by-minute time series data of the air conditioning adaptive control from June 6 to July 5, 2022 (a total of 43,200 sets of time series data) are converted into a 5-minute time series, totaling 8,640 samples. The input of the sample data is the load power P dev , Cabinet ambient temperature T env , cold channel temperature T inlet , hot channel temperature T outlet , output the air conditioning cooling temperature D. 70% of the data in the sample is used as the training set, 30% as the model validation set, and July 6 and July 7, when the indoor and outdoor temperatures are similar and stable, are selected to conduct a comparative test on the effect of the energy-saving optimization scheme.
[0143] When the load standby state of a single-cabinet data center is changed to a full load of 3Kw, the temperature control process based on the prediction model and its control effect are demonstrated. Figure 8It shows that on July 7, 2022, when the outdoor ambient temperature was 33°C, the adaptive control method began to execute when the load of the single-cabinet data center changed from standby to full load 3Kw. It can be concluded from the experimental diagram that the prediction model can quickly analyze the temperature setting value suitable for the current load heat dissipation, so that the micromodule control system can quickly send the air setting temperature to 21.81°C, which is equivalent to predicting the load heat dissipation in the future, matching the cooling in advance, and controlling the cabinet temperature to be stable within the boundary range of safe operation.
[0144] Using neural networks to model the temperature field of data centers and bring them into energy consumption optimization analysis can obtain more comprehensive analysis results. Figure 7 and Figure 8 It can be seen that compared with the adaptive control method, the temperature control algorithm based on the prediction model can quickly complete the temperature stabilization adjustment. As shown in the data comparison in Table 2, in the same test scenario, the adaptive control method takes 80 minutes to achieve hot and cold balance, while the temperature control method based on the prediction model can complete hot and cold balance within 40 minutes. By calculating the PUE of both within 5 to 80 minutes before stabilization and within 80 to 120 minutes after stabilization, the temperature control method based on the prediction model has improved by more than 2 percentage points, achieving the effect of rapid control and energy saving.
[0145] Table 1. Experimental results of single-cabinet data center air conditioning energy saving method based on new intelligent cooling module
[0146] Control Methods Outdoor temperature Stabilization time / min PUE between 5 and 80 PUE between 80 and 120 Adaptive control of supply air temperature 33℃ 80 1.44 1.46 Temperature control based on predictive model 33℃ 40 1.40 1.43
[0147] The above partial data does not represent all PUE optimization results. As the outdoor temperature decreases, the heat exchange efficiency of the air conditioning system increases, the system PUE also decreases, and the optimization effect becomes more obvious. Therefore, the data shown here is only used to further improve the energy saving effect of a single cabinet data center under the same conditions.
[0148] The embodiment of the present invention uses a large amount of measured data obtained when the energy-efficient optimal operating state is achieved under the adaptive control method to construct a temperature field model based on a neural network that fully considers the temperature field characteristics of the data center, and obtains the nonlinear relationship between the load power, hot channel temperature, cold channel temperature, cabinet ambient temperature and air conditioning cooling temperature. In the scenario where the load power fluctuates greatly and the refrigeration system requires a long adjustment time, a temperature control method based on a prediction model can be adopted, and the air conditioning cooling temperature predicted by the temperature field model is used for constant temperature control. Since the influence of the load fluctuation degree is fully considered, the supply air / return air temperature control module continuously seeks the best between the adaptive control method and the temperature control method based on the prediction model, and the coordinated control realizes the optimal energy efficiency of the refrigeration system. While quickly, accurately and effectively meeting the cabinet temperature requirements, energy consumption can be further reduced. Compared with the traditional artificial intelligence air conditioning energy-saving control model, a good refrigeration regulation effect is achieved with relatively few historical samples and a shorter learning time.
[0149] Based on the same inventive concept, an embodiment of the present invention provides a temperature control system for a single-cabinet data center, including: a temperature acquisition module, a power acquisition module and a temperature control module. The temperature acquisition module (i.e., the air inlet area temperature sensor 101, the air outlet area temperature sensor 102, and the ambient temperature sensor 103) is used to collect the cold channel temperature, the hot channel temperature, the running air conditioning cooling temperature, and the cabinet ambient temperature; wherein the running air conditioning cooling temperature can be provided by the monitoring management platform 104. The power acquisition module (which can be set on the monitoring management platform 104) is used to collect load power. The temperature control module (which can be set on the monitoring management platform 104) is used to execute the temperature control method of the single-cabinet data center. According to other embodiments of the present invention, the temperature control system of the single-cabinet data center also includes the training module, which is used to execute the training method of the prediction model, such as being set on the monitoring management platform 104.
[0150] The various embodiments of this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments or implementations. The same or similar parts between the various embodiments of the present invention can be referred to each other, and the implementation principles of the inventive concept and the technical effects produced can be referred to each other, which will not be repeated here. In the absence of conflict, the various embodiments or implementations of the present invention can be combined with each other.
[0151] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.
[0152] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.
[0153] Computer readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. Computer readable storage medium can include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.
[0154] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used in the embodiments of the present invention is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed in the embodiments of the present invention.
Claims
1. A training method for a prediction model for temperature control of a single-cabinet data center, comprising: The time series data of load power, cabinet ambient temperature, cold channel temperature, hot channel temperature and air conditioning cooling temperature at historical moments are obtained at preset time intervals and preset time windows as training samples; wherein the time series data is generated during the historical temperature control process of a single cabinet data center, the load power, the cabinet ambient temperature, the cold channel temperature and the hot channel temperature are feature data used as inputs of a neural network model, and the air conditioning cooling temperature is a label used as an expected output value; Iteratively optimizing the weights and bias parameters of the neural network model according to the data in the training sample until the error loss is lower than a preset error threshold or the number of iterations reaches a preset iteration threshold; The trained neural network model is used as a prediction model for temperature control of the single-cabinet data center.
2. The method according to claim 1, characterized in that The time series data of load power, cabinet ambient temperature, cold aisle temperature, hot aisle temperature and air conditioning cooling temperature at historical moments are obtained at preset time intervals and preset time windows as training samples, including: For the time series data of load power, cabinet ambient temperature, cold channel temperature, hot channel temperature and air conditioning cooling temperature at historical moments acquired at the first preset time interval and the first preset time window, resample the data at the second preset time interval and the second preset time window to construct a single resampled time series data; All resampled time series data are used as training samples; The first preset time interval is smaller than the second preset time interval, and the first preset time window is smaller than the second preset time window.
3. The method according to claim 1, characterized in that The historical temperature control process of the single-cabinet data center is a historical process of controlling the temperature of the single-cabinet data center by an adaptive control method.
4. The method according to claim 1, characterized in that The neural network model is an error back propagation neural network.
5. A temperature control method for a single cabinet data center, comprising: Obtaining the current load power, cabinet ambient temperature, cold channel temperature, and hot channel temperature of the single-cabinet data center; Using the prediction model obtained according to any one of the methods of claims 1 to 4, outputting the air conditioning cooling temperature at the next moment according to the load power, cabinet ambient temperature, cold channel temperature, and hot channel temperature at the current moment; The cooling demand is calculated based on the air-conditioning cooling temperature at the next moment and a cooling operation is performed.
6. A temperature control method for a single cabinet data center, comprising: Obtaining the current load power of the single-cabinet data center and calculating the fluctuation rate of the load power; If the fluctuation rate exceeds a preset fluctuation rate threshold, the temperature control method for a single-cabinet data center based on a neural network of claim 5 is executed; If the fluctuation rate does not exceed the preset fluctuation rate threshold, an adaptive control method is used to control the temperature of the single-cabinet data center.
7. The method according to claim 6, characterized in that In the process of executing the temperature control method of a single-cabinet data center based on a neural network, the method also includes: If the cold channel temperature at the next moment exceeds the preset cold channel temperature range, or the hot channel temperature at the next moment exceeds the preset hot channel temperature range, an adaptive control method is used to control the temperature of the single cabinet data center.
8. The method according to claim 6, characterized in that In the process of executing the temperature control method of a single-cabinet data center based on a neural network, the method also includes: If the temperature control method of the single-cabinet data center based on the neural network has been continuously executed for a time length exceeding a preset time length, an adaptive control method is used to control the temperature of the single-cabinet data center.
9. The method according to any one of claims 6 to 8, characterized in that: The method of controlling the temperature of the single-cabinet data center by using an adaptive control method includes: Obtaining the current cold channel temperature, hot channel temperature, and operating air conditioning cooling temperature of the single cabinet data center; Calculate the air conditioning cooling temperature at the next moment according to the preset cold channel temperature range, the preset hot channel temperature range, the load power at the current moment, the cold channel temperature, the hot channel temperature, and the running air conditioning cooling temperature; The cooling demand is calculated based on the air-conditioning cooling temperature at the next moment and a cooling operation is performed.
10. The method according to any one of claims 1 to 5 and 9, characterized in that: In the return air control mode, the air conditioning cooling temperature is the return air set temperature; or In the air supply control mode, the air conditioning cooling temperature is the air supply set temperature.
11. A temperature control system for a single cabinet data center, comprising: Temperature acquisition module, power acquisition module and temperature control module; The temperature acquisition module is used to collect the cold channel temperature, hot channel temperature, running air conditioning cooling temperature, and cabinet ambient temperature; The power collection module is used to collect load power; The temperature control module is used to execute the temperature control method as claimed in any one of claims 5 to 10.
12. A storage medium, characterized in that: The storage medium stores computer executable instructions, and when the computer executable instructions are loaded and executed by the processor, the steps of any method as claimed in claim 1 to 10 are implemented.
Citation Information
Cited By
Roadway local cooling control method and system based on deep reinforcement learning
CN119882879A
Method, device and system for controlling temperature and humidity in machine room
CN120447668A
Intelligent cabinet temperature control management system and equipment and use method
CN120560390A
