Model predictive control method and system for computer room temperature

Abstract

The present application relates to the field of data processing, and provides a model prediction control method and system for the temperature of a computer room, which obtains the control evaluation factor data of the computer room to be controlled; generates a first fused feature array, and loads it into multiple first task models to obtain the first task model extraction array output by each first task model; determines the completeness of the control evaluation factor data in each evaluation factor set in advance, and generates a primary data integrity array representation based on the completeness of each evaluation factor set and the feature array set for each evaluation factor set; determines the influence coefficient of each first task model based on the primary data integrity array representation, fuses each first task model extraction array based on the influence coefficient to obtain a primary blank influence fusion array, and obtains the temperature control parameter based on the merging result of the primary blank influence fusion array and each first task model extraction array. The present application improves the prediction accuracy of the temperature control parameter.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a model predictive control method and system for computer room temperature. Background Art

[0002] In the era of highly developed modern information technology, computer rooms are places where various computer systems, servers, etc. are stored together, and temperature control of their operating environment is crucial. Among traditional computer room temperature control methods, control based on fixed rules is more common. However, the actual heat generation in the computer room is dynamically affected by many factors, including the real-time load of the equipment, the increase or decrease of equipment in the computer room, and changes in the external ambient temperature. Fixed rules cannot adapt to these dynamic changes, which can easily lead to inaccurate temperature control, either causing energy waste (over-cooling or over-heating) or failing to meet the heat dissipation needs of the equipment.

[0003] In order to overcome the shortcomings of fixed rule control, some predictive control methods based on simple models have been proposed. These models may consider some factors that affect the temperature of the computer room, such as the simple linear relationship between equipment power and temperature, and then make a certain degree of temperature prediction based on historical data, and adjust the parameters of the temperature control equipment accordingly. However, these simple models often ignore many complex factors and their interrelationships. Moreover, these simple models lack effective response mechanisms when faced with incomplete data, which greatly reduces the accuracy of the prediction results. When building a computer room temperature control model, data integrity is a key issue. The data related to the computer room comes from a wide range of sources, including various computer room facility parameters, historical control data of different temperature control devices, and time series data of control operations. However, in the actual data collection process, due to equipment failure, data transmission interruption, imperfect collection system and other reasons, data is often missing or incomplete. For example, the historical control parameters of some temperature control equipment may be lost due to damage to the storage device, or when collecting computer room facility parameters, some parameters (such as thermal insulation performance indicators of specific areas) are not accurately obtained due to measurement difficulties. Traditional computer room temperature control models often perform poorly when dealing with incomplete data and are unable to accurately analyze and use existing data to predict temperature control parameters, thus affecting the effectiveness of computer room temperature control.

[0004] In summary, with the continuous development of information technology, the scale and complexity of computer rooms are increasing, and the requirements for computer room temperature control technology are becoming higher and higher. Existing computer room temperature control methods have many limitations in dealing with complex computer room environments, multiple influencing factors, and data incompleteness. Therefore, there is an urgent need for a new computer room temperature control technology that can fully consider the complex relationship between various factors in the computer room and can adaptively identify the blank state of the data, so as to accurately predict the temperature control parameters and improve the accuracy of computer room temperature control. Summary of the invention

[0005] The object of the present invention is to provide a model predictive control method and system for computer room temperature, and the present application is implemented as follows:

[0006] In a first aspect, the present application provides a model predictive control method for computer room temperature, the method comprising: obtaining control evaluation factor data of a computer room to be controlled, the control evaluation factor data comprising observation results corresponding to the computer room to be controlled and a plurality of reference indicators, respectively, the plurality of reference indicators comprising a plurality of computer room facility parameters, a plurality of temperature history control data and a plurality of temperature history control time series data, the temperature history control data comprising temperature control parameters of the computer room to be controlled on a plurality of temperature control devices, the temperature history control time series data representing temperature control device control time series data obtained by constructing a target number of controls of the computer room to be controlled on the plurality of temperature control devices before a current moment; performing feature representation on the observation results corresponding to each computer room facility parameter and the temperature history control data to obtain a first feature array, performing feature representation on the time series observation results corresponding to each temperature history control time series data to obtain a second feature array, and generating a first fused feature array based on each first feature array and each second feature array; The first fused feature array is loaded into a plurality of first task models respectively, and feature extraction is performed based on each first task model to obtain a first task model extraction array output by each first task model; the completeness of the control evaluation factor data in each evaluation factor set in advance is determined, and based on the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, a primary data integrity array representation of the control evaluation factor data is generated, and the primary data integrity array representation is used to identify the blank content of the observation results corresponding to the reference indicators in the control evaluation factor data; based on the primary data integrity array representation, an influence coefficient of each first task model is determined, and based on the influence coefficient, the extraction arrays of each first task model are fused to obtain a primary blank influence fusion array, and based on the merging result of the primary blank influence fusion array and the extraction arrays of each first task model, the temperature control parameters of the computer room to be controlled on the set temperature control equipment are obtained.

[0007] In a second aspect, the present application provides a computer system comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0008] Beneficial effects of the present invention: After obtaining the control evaluation factor data of the room to be controlled, the present application features the observation results corresponding to the facility parameters of each room and the temperature history control data to obtain a first feature array. In addition, the time series observation results corresponding to each temperature history control time series data are featured to obtain a second feature array. Then, a first fusion feature array is generated based on each first feature array and each second feature array. Based on this, feature extraction is performed based on each first task model to obtain a first task model extraction array corresponding to the first fusion feature array, and according to the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, a primary data integrity array representation of the control evaluation factor data is generated, and the blank state information of the observation result can be passed to the model. Then, based on the primary data integrity array representation, the influence coefficient of each first task model is determined, and the extraction array of each first task model is fused based on the influence coefficient to obtain a primary blank influence fusion array, and the temperature control parameters of the room to be controlled on the set temperature control device are obtained based on the merging result of the primary blank influence fusion array and the extraction array of each task model, so that the model can adaptively identify the data blank state, so as to accurately help the accurate prediction of the temperature control parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flow chart of a model predictive control method for computer room temperature provided in an embodiment of the present application.

[0010] Figure 2 It is a schematic diagram of the composition of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0011] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method part of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0012] The execution subject of the model prediction control method of the room temperature in the embodiment of the present application is a computer system, including but not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes:

[0013] Step S100: Obtain control evaluation factor data of the computer room to be controlled, the control evaluation factor data include observation results corresponding to the computer room to be controlled and multiple reference indicators respectively, the multiple reference indicators include multiple computer room facility parameters, multiple temperature history control data and multiple temperature history control timing data, the temperature history control data include temperature control parameters of the computer room to be controlled on multiple temperature control devices, and the temperature history control timing data represents the temperature control device control timing data obtained based on the target number of controls of the computer room to be controlled on multiple temperature control devices before the current moment.

[0014] In an embodiment of the present application, the reference indicators include multiple computer room facility parameters, multiple temperature history control data, and multiple temperature history control timing data.

[0015] The parameters of the computer room facilities are the parameters that describe the characteristics of the facilities inside the computer room, such as the number of servers in the computer room, the power of the servers, the size of the computer room, the performance of the thermal insulation materials in the computer room, etc.

[0016] The temperature history control data includes the temperature control parameters of the control room on multiple temperature control devices. Temperature control devices can be air conditioners, fans and other devices used to adjust the temperature in the room. The temperature control parameters are some temperature control-related values ​​set by these devices during past operation. For example, the set temperature of the air conditioner is 20 degrees Celsius, the cooling power is set to 50%, and the fan speed is set to 1500 revolutions per minute. For air conditioners, different time periods may have different temperature settings, and these historical setting values ​​are part of the temperature history control data.

[0017] The temperature history control time series data represents the temperature control device control time series data obtained based on the control construction of the target number of times on multiple temperature control devices in the control room before the current moment. For example, assuming the target number is 5 times, in terms of the dimension of cooling trigger, the temperature control device control time series data is a 5-dimensional vector. If the three most recent cooling triggers were completed by three different air-conditioning devices, then the device codes of these three devices will be recorded in this vector (assuming the device codes are 101, 102, and 103 respectively). Since there are less than five devices, the remaining two elements are filled with 0 to form a 5-dimensional vector (101, 102, 103, 0, 0). For another example, in the dimension of fan speed increase trigger, assuming the control construction with the target number of 5 times, if the four most recent fan speed increase triggers were completed by four fan devices, and the device codes are 201, 202, 203, and 204 respectively, then the corresponding vector is (201, 202, 203, 204, 0). The device code here is a code used to uniquely identify each temperature control device, which facilitates the computer system to identify and process data. In addition to the two dimensions of cooling trigger and fan speed trigger, there can be other dimensions, such as the activation of the dehumidification function, the activation of the heating function (heating may be required to maintain temperature stability in some special computer room environments), etc. Taking the activation of the dehumidification function as an example, the corresponding temperature control devices are listed in chronological order. If the dehumidification function has been activated twice recently, they were completed by devices with device codes 301 and 302 respectively, and the constructed vector is (301, 302, 0, 0, 0). In the dimension of heating function activation, if the heating has been activated once recently and completed by a device with a device code of 401, the vector is (401, 0, 0, 0, 0).

[0018] Step S200: Characterize the observation results corresponding to each computer room facility parameter and temperature history control data to obtain a first feature array, characterize the time series observation results corresponding to each temperature history control time series data to obtain a second feature array, and generate a first fused feature array based on each first feature array and each second feature array.

[0019] In step S200 of the embodiment of the present application, the computer system features the observation results corresponding to each computer room facility parameter and the temperature history control data to obtain a first feature array. For the computer room facility parameters, taking the server power as an example, assuming that the server power takes values ​​within a certain range, the computer system can convert it into a feature representation through a certain numerical mapping method. For example, the value range of the server power is 100 watts to 1000 watts, and the computer system can map 100 watts to feature value 1, 200 watts to feature value 2, and so on. If there are multiple computer room facility parameters, such as the number of servers, the size of the computer room space, etc., the computer system will convert the observation results corresponding to these parameters into feature values ​​according to their respective pre-set mapping rules or feature extraction methods, and then combine these feature values ​​into part of the first feature array.

[0020] For the temperature control parameters in the temperature history control data, such as the set temperature and cooling power of the air conditioner. The set temperature value range is assumed to be 15-30 degrees Celsius. The computer system can map 15 degrees Celsius to a certain characteristic value, such as 1. For every increase of 1 degree Celsius, the characteristic value increases by 1. If the cooling power value range is 0-100%, 0% can be mapped to the characteristic value 1, and for every increase of 10%, the characteristic value increases by 1. These mapped characteristic values ​​and the characteristic values ​​of the equipment room facility parameters together form the first characteristic array.

[0021] Next, the computer system features the time series observation results corresponding to each temperature history control time series data to obtain the second feature array. Taking the previously mentioned refrigeration-triggered temperature control device control time series data as an example, assume that the 5-dimensional vector (101, 102, 103, 0, 0) represents the most recent refrigeration trigger device code sequence. The computer system can perform one-hot encoding on each element in this vector. For a device with a device code of 101, if there are a total of 1000 possible device codes, then under one-hot encoding, the encoding vector corresponding to this device code is a 1000-dimensional vector with 1 at the 101st position and 0 at the remaining positions. Similarly, the device codes 102, 103, etc. are also encoded in this way, and then these encoding vectors are combined as part of the second feature array. A similar one-hot encoding method is also used for the temperature control device control time series data (201, 202, 203, 204, 0) triggered by the fan speed increase, and the obtained encoding vector is also added to the second feature array. If there are temperature control equipment control timing data in dimensions such as dehumidification function startup and heating function startup, the features are all represented in the same way to eventually form a complete second feature array.

[0022] Finally, the computer system generates a first fused feature array based on each first feature array and each second feature array. This process can be done in a variety of ways, such as directly splicing the first feature array and the second feature array in order. Assuming that the first feature array is a 100-dimensional array and the second feature array is a 200-dimensional array, the computer system can connect the two arrays into a 300-dimensional array as the first fused feature array. Alternatively, the computer system can also use a weighted combination method to set weights for the first feature array and the second feature array, for example, the weight of the first feature array is 0.6 and the weight of the second feature array is 0.4. Then multiply each element in the first feature array by 0.6, multiply each element in the second feature array by 0.4, and then add the corresponding elements to obtain a new fused array as the first fused feature array.

[0023] By converting the room facility parameters and temperature history control data into the first feature array, the computer system represents these data in a form that is more suitable for model processing, which is convenient for subsequent feature extraction and analysis. The second feature array is obtained by feature representation of the temperature history control time series data, which can effectively capture the laws and characteristics of temperature control in the time series. Generating the first fusion feature array integrates these different aspects of feature information together, so that the subsequent model can comprehensively consider various relevant information of the room, so as to more accurately predict the room temperature control parameters. This helps to improve the accuracy and efficiency of the temperature control of the entire room, ensure the normal operation environment of the equipment in the room, and reduce problems such as equipment failure and performance degradation caused by abnormal temperature.

[0024] Step S300: loading the first fusion feature array into a plurality of first task models respectively, and performing feature extraction based on each first task model to obtain a first task model extraction array output by each first task model.

[0025] In an embodiment of the present application, the first task model (Expert Network) is an expert model specially designed to process specific types of tasks or data features. These models are set to have good recognition and processing capabilities for specific data patterns or feature relationships when they are constructed. For example, there may be a first task model specifically for the relationship between computer room facility parameters and temperature control, and a first task model for temperature history control data patterns, etc. The computer system loads the first fused feature array into these different first task models. Assume that the first fused feature array is an array with a dimension of 500, which contains the feature information of the computer room facility parameters, temperature history control data, and temperature history control timing data that have been processed and integrated before.

[0026] Take a first task model for the relationship between room facility parameters and temperature control as an example. This model may have a multi-layer neural network structure, such as an input layer, a hidden layer, and an output layer. When the computer system loads the 500-dimensional first fused feature array into the input layer of this model, the number of neurons in the input layer will correspond to the dimension of the first fused feature array, that is, there are 500 neurons to receive these feature data.

[0027] In the hidden layer of this model, neurons process and transmit data through pre-set weights and activation functions. For example, neurons in the hidden layer may use the linear rectification function (ReLU) as the activation function. Assuming that there are 100 neurons in the hidden layer, each neuron will perform a weighted summation of the 500 input feature data according to their respective weights, and perform a nonlinear transformation through the ReLU function. This nonlinear transformation can help the model better capture the complex relationships in the data. For the output layer, according to the design purpose of this first task model, it will output the feature extraction results related to the task to form the first task model extraction array. If the purpose of this model is to extract the characteristics of the influence of the room facility parameters on temperature control, then the output first task model extraction array may be a 10-dimensional array, each dimension represents a specific feature relationship, such as the weight of the influence of server power on temperature control, the coefficient of influence of the room space size on temperature regulation efficiency, etc.

[0028] Taking another first task model for the temperature history control data pattern as an example, its internal structure also has an input layer, a hidden layer and an output layer. After the input layer receives the 500-dimensional first fusion feature array, similar data processing based on weights and activation functions (different activation functions may be used, such as the Sigmoid function) is performed in the hidden layer. The number of neurons and the connection method of the hidden layer depend on the design of the model, for example, there are 80 neurons. After processing by the hidden layer, a first task model extraction array related to the feature extraction of the temperature history control data is output in the output layer. If this model focuses on the impact of the historical fluctuation pattern of the temperature control parameters on the current temperature control, then the output array may be a 15-dimensional array, which contains information such as the impact factors of the fluctuation of the temperature control parameters in different historical periods on the current temperature setting.

[0029] By loading the first fusion feature array into multiple such first task models, the computer system can extract features from the data from different angles. Based on its own structure and function, each first task model mines the feature information related to a specific task in the first fusion feature array, thereby obtaining its own first task model extraction array. These extraction arrays contain deep feature information on the data related to the temperature control of the computer room from different aspects, laying the foundation for further integration of this information to accurately predict the temperature control parameters of the computer room. This application method of multi-task models can make full use of the sensitivity of different models to different data features, improve the accuracy and reliability of the entire computer room temperature prediction control system, and ensure that the temperature in the computer room can be effectively controlled to meet the strict requirements of the equipment in the computer room for the operating environment temperature.

[0030] Step S400: Determine the completeness of the control evaluation factor data in each pre-set evaluation factor set, and generate a primary data integrity array representation of the control evaluation factor data based on the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set. The primary data integrity array representation is used to identify the blank content of the observation results corresponding to the reference indicators in the control evaluation factor data.

[0031] In the embodiment of the present application, the evaluation element set is a set of control evaluation element data classified and combined according to certain rules or logic. For example, an evaluation element set may be a set of elements related to the parameters of the computer room facilities, including the number of servers, server power, the size of the computer room space, etc.; another evaluation element set may be a set of elements related to the temperature history control data, such as the historical set temperature of the air conditioner, the cooling power, the historical speed of the fan, etc.; there may also be an evaluation element set specifically for the temperature history control time series data, such as the time series data of the cooling trigger, the time series data of the fan speed trigger, and other time series data sets of different dimensions. The computer system determines the completeness of the control evaluation element data in each evaluation element set. Taking the computer room facility parameter evaluation element set as an example, it is assumed that this set contains 5 elements, namely the number of servers, server power, the size of the computer room space, the performance of the thermal insulation material, and the number of ventilation equipment. The computer system will check the observation results of these 5 elements in the control evaluation element data for the computer room. If it is found that there is no observation result (ie, data missing) for the thermal insulation material performance element, then the number of blank evaluation elements corresponding to the blank observation result is 1. The total number of evaluation factors in this evaluation factor set is 5, and the total number of complete evaluation factors is 5-1=4. Divide the total number of complete evaluation factors by the total number of evaluation factors, that is, 4÷5=0.8, and the completeness of the equipment room facility parameter evaluation factor set is 0.8.

[0032] Let's look at the temperature history control data evaluation element set. Assume that it contains 4 elements: the historical set temperature of air conditioner 1, the cooling power of air conditioner 1, the historical speed of fan 1, and the historical set temperature of air conditioner 2. If there is no record of the historical set temperature of air conditioner 2 in the control evaluation element data, the number of blank evaluation elements is 1, the total number of evaluation elements is 4, and the total number of complete evaluation elements is 3. The completeness of the evaluation element set is 3÷4=0.75.

[0033] For the temperature history control time series data evaluation element set, for example, in the cooling trigger time series data dimension, it is assumed that it contains three elements: the equipment code of the last three cooling triggers, the time interval of the cooling triggers, and the ambient temperature when the cooling is triggered. If there is no observation result of the ambient temperature when the cooling is triggered, the number of blank evaluation elements is 1, the total number of evaluation elements is 3, and the total number of complete evaluation elements is 2, and its completeness is 2÷3≈0.67.

[0034] Next, the computer system generates a primary data integrity array representation of the control evaluation factor data based on the integrity of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set. Each evaluation factor set has a pre-set feature array, which is set according to the characteristics of the evaluation factor set and subsequent processing requirements. For example, for the computer room facility parameter evaluation factor set, the set feature array may be an array such as [2, 3, 1, 4, 2].

[0035] Multiply the completeness of each evaluation feature set by its set feature array. For the room facility parameter evaluation feature set, the completeness is 0.8. Multiply it by each element in the feature array [2, 3, 1, 4, 2] to obtain [2×0.8, 3×0.8, 1×0.8, 4×0.8, 2×0.8]=[1.6, 2.4, 0.8, 3.2, 1.6]. For the temperature history control data evaluation feature set, the completeness is 0.75. Assuming that its feature array is [3, 1, 2, 1], after multiplication, we get [3×0.75, 1×0.75, 2×0.75, 1×0.75]=[2.25, 0.75, 1.5, 0.75]. For the temperature history control timing data evaluation feature set, the completeness is 0.67. Assuming its feature array is [1, 2, 3], after multiplication, we get [1×0.67, 2×0.67, 3×0.67]=[0.67, 1.34, 2.01].

[0036] Then, the computer system combines the adjustment feature arrays corresponding to each evaluation factor set to obtain a primary data integrity array representation. Assuming that the adjustment feature arrays corresponding to the above three evaluation factor sets are combined, the primary data integrity array representation obtained may be [1.6, 2.4, 0.8, 3.2, 1.6, 2.25, 0.75, 1.5, 0.75,0.67, 1.34, 2.01]. This array can identify the blank content of the observation results corresponding to the reference indicators in the control evaluation factor data. Through the operation and combination of integrity and feature arrays, the data integrity of different evaluation factor sets is reflected in this primary data integrity array representation in a quantitative form.

[0037] This primary data integrity array representation enables subsequent processing steps to clearly understand the data integrity of each evaluation element set, so that the blank state of the data can be effectively considered and processed during the model processing. For example, when the data is subsequently processed based on the first task model, this array representation can be used as a reference information to assist the model in adjusting the processing method for different data to adapt to the situation of incomplete data, improve the accuracy and reliability of the temperature prediction control of the entire computer room, and ensure that the equipment in the computer room operates stably under a suitable temperature environment.

[0038] Step S500: Determine the influence coefficient of each first task model based on the primary data integrity array representation, fuse the extraction arrays of each first task model based on the influence coefficient to obtain a primary blank influence fusion array, and obtain the temperature control parameters of the computer room to be controlled on the set temperature control device based on the merging result of the primary blank influence fusion array and the extraction arrays of each first task model.

[0039] In an embodiment of the present application, the computer system determines the influence coefficient of each first task model based on the primary data integrity array representation, that is, a weight information. The primary data integrity array representation contains the integrity information of the control evaluation factor data in each evaluation factor set, which reflects the integrity of the data. The first task model is the model used to extract features from the first fused feature array in the previous step. For example, suppose there are three first task models, namely model A, model B and model C. Different elements in the primary data integrity array representation correspond to data integrity related information of different evaluation factor sets.

[0040] Assume that the primary data integrity array is represented as [0.8, 0.7, 0.6], where each value may correspond to the data integrity of the evaluation factor set related to the three first task models. The computer system determines the influence coefficient according to a preset rule. An exemplary rule is to use the data integrity value directly as the influence coefficient, then the influence coefficient of model A is 0.8, the influence coefficient of model B is 0.7, and the influence coefficient of model C is 0.6. Another possible rule is to perform some transformation on the data integrity value to obtain the influence coefficient, such as normalizing each data integrity value by dividing it by their sum. The sum is 0.8+0.7+0.6=2.1, then the influence coefficient of model A is 0.8 / 2.1≈0.38, the influence coefficient of model B is 0.7 / 2.1≈0.33, and the influence coefficient of model C is 0.6 / 2.1≈0.29.

[0041] Next, the computer system fuses the first task model extraction arrays based on the determined influence coefficients to obtain a primary blank influence fusion array. The first task model extraction array is an array output by each first task model after feature extraction on the first fusion feature array. For example, the first task model extraction array output by model A is [1, 2, 3], the first task model extraction array output by model B is [4, 5, 6], and the first task model extraction array output by model C is [7,8, 9].

[0042] If the influence coefficient is used directly as described above (the influence coefficient of model A is 0.8, the influence coefficient of model B is 0.7, and the influence coefficient of model C is 0.6), the computer system performs weighted summation to fuse these arrays. For the first element of the fused array, the calculation method is: 1×0.8+4×0.7+7×0.6=0.8+2.8+4.2=7.8; for the second element: 2×0.8+5×0.7+8×0.6=1.6+3.5+4.8=9.9; for the third element: 3×0.8+6×0.7+9×0.6=2.4+4.2+5.4=12. So the primary blank influence fused array obtained is [7.8, 9.9, 12]. If the normalized influence coefficient is used (model A is 0.38, model B is 0.33, and model C is 0.29), then for the first element: 1×0.38+4×0.33+7×0.29=0.38+1.32+2.03=3.73; for the second element: 2×0.38+5×0.33+8×0.29=0.76+1.65+2.32=4.73; for the third element: 3×0.38+6×0.33+9×0.29=1.14+1.98+2.61=5.73. The primary blank influence fusion array obtained at this time is [3.73, 4.73, 5.73].

[0043] Then, the computer system obtains the temperature control parameters of the room to be controlled on the set temperature control device based on the merging result of the primary blank influence fusion array and the first task model extraction arrays. The merging result here can be a combination of various forms. One way is to directly splice the primary blank influence fusion array and the first task model extraction arrays. For example, according to the first influence coefficient calculation method above, the primary blank influence fusion array [7.8, 9.9, 12], the first task model extraction array of model A [1, 2, 3], model B [4, 5, 6], and model C [7, 8, 9], the spliced ​​array may be [7.8, 9.9, 12, 1, 2, 3, 4, 5, 6, 7, 8, 9].

[0044] This merged result contains feature information extracted from different first task models and fusion information after considering data integrity. The computer system determines the temperature control parameters of the room to be controlled on the set temperature control equipment based on this merged result. This process may involve complex algorithms and models. For example, there may be a predefined mapping relationship that converts the elements in the merged array into temperature control parameters through some mathematical operation. Assume that this mapping relationship is a multivariate linear regression model, taking the elements in the merged array as independent variables, and calculating the value of the temperature control parameter through the regression equation.

[0045] Alternatively, there is a rule-based system that determines the temperature control parameters according to the values ​​of the elements in the merged array falling into different intervals. For example, if the value of a key element in the merged array is within a certain interval, the temperature of the temperature control device (such as air conditioner) is set to 20 degrees Celsius and the cooling power is set to 50%; if it is in another interval, the temperature is set to 22 degrees Celsius and the cooling power is set to 40%, etc.

[0046] This method of determining the influence coefficient based on data integrity, extracting the array through the fusion model, and determining the temperature control parameters based on the merging results fully considers the impact of data integrity on the model output. Since in actual situations, the control evaluation factor data may be incomplete, this method can make the model better adapt to the incompleteness of the data and improve the accuracy of the prediction of the temperature control parameters of the computer room. Accurate temperature control parameters are crucial to the normal operation of the equipment in the computer room, which can ensure that the temperature in the computer room is kept within the appropriate range to avoid adverse effects on equipment performance and life due to excessively high or low temperatures.

[0047] As an implementation mode, in step S500, the temperature control parameters of the room to be controlled on the set temperature control device are obtained according to the merging result of the primary blank impact fusion array and the first task model extraction arrays, including:

[0048] Step S510: combining the primary blank impact fusion array and each first task model extraction array to obtain a second fusion feature array;

[0049] Step S520: loading the second fused feature array or the third fused feature array generated based on the second fused feature array into the deep dense network to obtain a mapping result of the deep dense network;

[0050] Step S530: The mapping results are loaded into the S-type nonlinear transformation component, the identity transformation component and the smooth ReLU transformation component respectively, and the confidence of triggering the cooling control of the computer room to be controlled, the expectation of the temperature control parameters of the computer room to be controlled and the variance deviation of the temperature control parameters of the computer room to be controlled are inferred respectively;

[0051] Step S540: Determine and set the temperature control parameters on the temperature control device based on the confidence, expectation and variance deviation.

[0052] When the computer system executes step S510, the previously obtained primary blank influence fusion array and each first task model extraction array are combined. For example, assuming that the primary blank influence fusion array is [3.73, 4.73, 5.73] (this array is obtained by fusing each first task model extraction array based on the previously determined influence coefficient), assuming that there are three first task models, their extraction arrays are: the first task model extraction array of model A is [1, 2, 3], that of model B is [4, 5, 6], and that of model C is [7, 8, 9].

[0053] There are many ways to combine, one of which is direct splicing. The computer system will splice these arrays together in order to obtain the second fused feature array [3.73, 4.73, 5.73, 1, 2, 3, 4, 5, 6, 7, 8, 9]. This combination method can integrate the information after fusion considering data integrity with the original feature information extracted by each first task model. The purpose of doing this is to be able to comprehensively utilize this information in subsequent processing to avoid information loss, so as to more comprehensively consider the impact of various factors on the temperature control parameters of the computer room.

[0054] In step S520, the computer system determines whether to load the second fused feature array directly into the deep dense network, or to generate a third fused feature array based on the second fused feature array and then load it.

[0055] A Dense Network is a fully connected neural network structure, where neurons in each layer are connected to all neurons in the next layer. Assuming that the second fused feature array is [3.73, 4.73, 5.73, 1, 2, 3,4, 5, 6, 7, 8, 9], when the computer system loads this array into the Dense Network, the number of neurons in the input layer of the Dense Network needs to match the dimension of the second fused feature array, that is, there are 12 input neurons to receive this data.

[0056] The deep dense network contains multiple hidden layers, and the number of neurons and activation function of each hidden layer are pre-set. For example, the first hidden layer may have 20 neurons, and the activation function is ReLU (rectified linear function). When the input data enters this hidden layer, each neuron will perform a weighted summation of the 12 input feature data according to their respective weights, and perform a nonlinear transformation through the ReLU function. This process will be repeated in each hidden layer.

[0057] Assume that the deep dense network has three hidden layers, the second hidden layer has 15 neurons, and the activation function is the Sigmoid function; the third hidden layer has 10 neurons, and the activation function is the Tanh function. After this series of processing, a mapping result will be obtained in the output layer. This mapping result is the output of the deep dense network after complex feature transformation and mapping of the second fused feature array of the input. It contains the deep dense network's abstract representation of the input data features, which may be a low-dimensional vector, such as [0.5, 0.3, 0.1]. Each element of this vector reflects some feature information related to the temperature control of the computer room learned by the deep dense network from the input data.

[0058] If the third fused feature array is generated based on the second fused feature array and then loaded into the deep dense network, some additional processing of the second fused feature array may be involved. For example, the elements in the second fused feature array are normalized, the value of each element is mapped to between 0 and 1, and then combined into the third fused feature array. Assume that the array obtained after normalization is [0.2, 0.3, 0.4, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7,0.8, 0.9]. Loading this third fused feature array into the deep dense network is similar to directly loading the second fused feature array, but because the data has been normalized, it may make the learning process of the deep dense network more stable and effective.

[0059] In step S530, after obtaining the mapping result of the deep dense network (for example, [0.5, 0.3, 0.1]), the computer system loads the result into different transformation components respectively.

[0060] The S-type nonlinear transformation component uses the Sigmoid function to transform the input data. The formula of the Sigmoid function is , where x is the input data (i.e., the element in the mapping result of the deep dense network). For example, for the first element 0.5 in the mapping result, after the Sigmoid function transformation, we get This result is interpreted as the confidence level of the room to be controlled triggering cooling control. If this value is close to 1, it means that the computer system has a high confidence that the room needs to trigger cooling control; if it is close to 0, it means that the confidence level is very low.

[0061] The identity component directly outputs the input data, that is, it directly outputs the elements in the mapping result of the deep dense network without any transformation. For example, for the second element 0.3 in the mapping result, the identity component directly outputs 0.3, which is regarded as the expected value of the temperature control parameter of the room to be controlled. This expectation reflects an expected value of the room temperature control parameter under the current data and model, such as the temperature setting value or the expected value of the temperature control device power.

[0062] The Soft ReLU Component uses the smooth ReLU function to transform the input data. The smooth ReLU function is a smooth approximation of the ReLU function. For example, for the third element 0.1 in the mapping result, a new value (assuming it is 0.05) is obtained after the smooth ReLU function transformation. This value is interpreted as the variance deviation of the temperature control parameter of the room to be controlled. The variance deviation reflects the fluctuation range or uncertainty of the temperature control parameter. A smaller variance deviation indicates that the temperature control parameter is relatively stable, and a larger variance deviation indicates that the temperature control parameter may fluctuate greatly.

[0063] In step S540, the computer system determines the temperature control parameters on the temperature control device based on the confidence, expectation and variance deviation obtained previously. Assume that the confidence obtained from the S-type nonlinear transformation component is 0.62, the expectation obtained from the identity transformation component is 0.3, and the variance deviation obtained from the smooth ReLU transformation component is 0.05. If the confidence 0.62 is higher than a preset threshold (for example, 0.5), it means that there is a high possibility that the computer room needs to be cooled. The computer system will determine the specific temperature control parameters based on the expectation 0.3 and the variance deviation 0.05.

[0064] For example, if the expectation is that 0.3 represents the adjustment of the temperature setting value relative to the current temperature (assuming the current temperature is 25 degrees Celsius), then the initially determined temperature setting value is 25-0.3=24.7 degrees Celsius. Then, considering the variance deviation of 0.05, the computer system may add a certain adjustment range based on this initially determined temperature value. If the variance deviation is set to represent the upper and lower floating range, the final temperature control parameter may be between 24.7-0.05 and 24.7+ 0.05, that is, between 24.65 and 24.75 degrees Celsius. This temperature control parameter will be applied to the set temperature control device (such as air conditioner) to achieve the control of the room temperature. If the confidence is lower than the threshold, it means that there is little need for cooling control. The computer system may maintain the current temperature control parameter or make fine adjustments according to other rules. For example, if the current cooling power of the air conditioner is 30%, since the confidence is low and cooling is not required, the computer system may adjust the cooling power to 25% to save energy while maintaining the relative stability of the room temperature. This method of determining temperature control parameters based on the results obtained from different transformation components fully considers the complexity and uncertainty of temperature control in the computer room. Confidence reflects the reliability of the decision, expectation gives a rough control target, and variance deviation takes into account the fluctuation factors in the actual situation. Combining these three aspects can more accurately determine the temperature control parameters suitable for the computer room environment, thereby ensuring that the equipment in the computer room operates stably in a suitable temperature environment and improving the performance and service life of the equipment.

[0065] As an implementation manner, in step S520, before loading the second fused feature array or the third fused feature array generated based on the second fused feature array into the deep dense network, the method may further include:

[0066] Step S511: loading the second fused feature array into a plurality of second task models respectively, and performing feature extraction based on each second task model to obtain a second task model extraction array output by each second task model;

[0067] Step S512: Based on the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, a secondary data integrity array representation of the control evaluation factor data is generated, wherein the secondary data integrity array representation is used to identify blank contents of the indicator combination between the reference indicators in the control evaluation factor data;

[0068] Step S513: determining the influence coefficient of each second task model based on the secondary data integrity array representation, and fusing each second task model extraction array based on the influence coefficient to obtain a secondary blank influence fusion array;

[0069] Step S514: Combine the secondary blank influence fusion array and the primary blank influence fusion array to obtain a third fusion feature array.

[0070] When the computer system executes step S511, it must first clarify the structure and content of the second fused feature array. Assume that the second fused feature array is obtained by combining the previous steps, for example, [3.73, 4.73, 5.73, 1, 2, 3,4, 5, 6, 7, 8, 9].

[0071] The second task model is a model specifically used to further extract features from the second fused feature array. These models may be different from the previous first task models in function and structure. They are designed to mine different types of feature information from the second fused feature array. For example, suppose there are three second task models, namely model D, model E and model F. When the computer system loads the second fused feature array into model D, the input layer of model D receives this array. Model D may have a specific network structure, for example, the number of neurons in its input layer is the same as the dimension of the second fused feature array, that is, 12 neurons. Model D contains a hidden layer. Assume that the hidden layer has 10 neurons and the activation function is ReLU. When the data enters the hidden layer, each neuron will perform a weighted summation of the 12 input feature data according to their respective weights, and perform a nonlinear transformation through the ReLU function. After being processed by the hidden layer, the feature information related to the task is output at the output layer to form the second task model extraction array output by model D. Assume that this array is [0.1, 0.2, 0.3]. Each element here represents a specific feature relationship extracted by model D from the second fused feature array, which may be related to different aspects of temperature control in the computer room. For example, the first element may be related to the impact of server heat dissipation on temperature control, the second element may be related to the impact of the ventilation structure of the computer room on temperature, and the third element may be related to the impact of the collaborative working relationship of temperature control equipment on temperature control.

[0072] Similarly, the second fused feature array is loaded into model E and model F. Model E may have a different network structure, for example, the input layer also has 12 neurons, the hidden layer has 8 neurons, and the activation function is the Sigmoid function. After processing, the second task model extraction array output by model E may be [0.4, 0.5, 0.6]. The input layer of model F also has 12 neurons, the hidden layer has 12 neurons, and the activation function is the Tanh function. The second task model extraction array output by model F may be [0.7, 0.8, 0.9]. These different second task model extraction arrays reflect the feature information in the second fused feature array from different angles, providing more information sources for subsequent further fusion and processing.

[0073] In step S512, the computer system once again considers the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set. Assume that there were three evaluation factor sets before, namely the computer room facility parameter evaluation factor set, the temperature history control data evaluation factor set and the temperature history control timing data evaluation factor set. For the computer room facility parameter evaluation factor set, its completeness is 0.8 (as calculated before), and it is assumed that the feature array set for it is [2, 3, 1, 4, 2]; the temperature history control data evaluation factor set has a completeness of 0.75, and a feature array of [3, 1, 2,1]; the temperature history control timing data evaluation factor set has a completeness of 0.67, and a feature array of [1, 2, 3].

[0074] Determine the completeness of the indicator combination and the feature array for each pair of evaluation element sets, taking the pair of the equipment room facility parameter evaluation element set and the temperature history control data evaluation element set as an example. The computer system first determines the completeness of the indicator combination of this pair of evaluation element sets. According to the previous rules, it is the product of the completeness of the two evaluation element sets, that is, 0.8×0.75=0.6. Then determine the feature array of this pair of evaluation element sets, which is the dot product of the feature arrays of the two evaluation element sets. The calculation method is: ; Take the pair of the equipment room facility parameter evaluation element set and the temperature history control time series data evaluation element set as an example. The indicator combination completeness is 0.8×0.67=0.536. The dot product of the feature array is calculated as: ; Finally, for the pair of temperature history control data evaluation element set and temperature history control time series data evaluation element set, the indicator combination completeness is 0.75×0.67=0.5025. The dot product of the feature array is calculated as: . Then, determine the secondary data integrity array representation of each pair of evaluation element sets and combine them. For the pair of the computer room facility parameter evaluation element set and the temperature history control data evaluation element set, the computer system multiplies the indicator combination integrity 0.6 and the calculation result 15 of the feature array to obtain a part of the secondary data integrity array representation of this pair, that is, 0.6×15=9. For the pair of the computer room facility parameter evaluation element set and the temperature history control timing data evaluation element set, 0.536×11=5.896. For the pair of the temperature history control data evaluation element set and the temperature history control timing data evaluation element set, 0.5025×11=5.5275.

[0075] The computer system combines these results to obtain a secondary data integrity array representation of the control evaluation factor data, such as [9, 5.896, 5.5275]. This array representation can identify the blank content of the indicator combination between the reference indicators in the control evaluation factor data. For example, if a certain value is low, it may mean that there are many blank indicator combinations between the corresponding two evaluation factor sets, that is, the data of some key indicator combinations are missing or incomplete, which will affect the subsequent accurate prediction of the temperature control parameters of the computer room.

[0076] In step S513, the computer system determines the influence coefficient of each second task model (model D, model E and model F) according to the obtained secondary data integrity array representation [9, 5.896, 5.5275].

[0077] One method to determine the influence coefficient is to first normalize the elements in the secondary data integrity array representation. The sum of the elements is calculated to be 9+5.896+5.5275=20.4235. Then, the influence coefficient corresponding to model D is 9 / 20.4235≈0.44, the influence coefficient corresponding to model E is 5.896 / 20.4235≈0.29, and the influence coefficient corresponding to model F is 5.5275 / 20.4235≈0.27.

[0078] Next, the computer system fuses the second task model extraction arrays based on these influence coefficients to obtain a secondary blank influence fusion array. Assume that the second task model extraction array output by model D is [0.1, 0.2, 0.3], the output by model E is [0.4, 0.5, 0.6], and the output by model F is [0.7, 0.8, 0.9]. For the first element of the fused array, the calculation method is: 0.1×0.44+0.4×0.29+0.7×0.27=0.044+0.116+0.189=0.349; for the second element: 0.2×0.44+0.5×0.29+0.8×0.27=0.088+0.145+0.216=0.449; for the third element: 0.3×0.44+0.6×0.29+0.9×0.27=0.132+0.174+0.243=0.549. So the obtained secondary blank influence fusion array is [0.349, 0.449, 0.549]. This secondary blank influence fusion array takes into account the data integrity of the combination of indicators between the second task model extraction arrays and the reference indicators, and can more accurately reflect the comprehensive characteristics of the data in subsequent processing.

[0079] In step S514, the computer system combines the secondary blank effect fusion array and the primary blank effect fusion array obtained previously. Assume that the primary blank effect fusion array is [3.73, 4.73, 5.73] and the secondary blank effect fusion array is [0.349, 0.449, 0.549].

[0080] The combination method can be direct concatenation, and the third fused feature array is obtained as [3.73, 4.73, 5.73, 0.349, 0.449, 0.549]. This combined third fused feature array integrates the information after considering the impact of data integrity at different levels, and provides a more comprehensive and accurate input reflecting data features for subsequent loading into the deep dense network (such as step S520), which helps to improve the accuracy and reliability of the prediction of the temperature control parameters of the entire computer room.

[0081] As an implementation manner, in step S200, generating a first fused feature array according to each first feature array and each second feature array includes:

[0082] Step S210: combining the first feature arrays into a fourth fused feature array, and combining the second feature arrays into a fifth fused feature array;

[0083] Step S220: determining the indicator combination relationship of the fourth fusion feature array based on the plurality of indicator combination models, and obtaining the first array representation output by each indicator combination model;

[0084] Step S230: loading the fifth fused feature array into the sequence processing recursive model to obtain a second array representation of the sequence processing recursive model output;

[0085] Step S240: Combine the first array representation and the second array representation into a first fused feature array.

[0086] In step S210, after acquiring each first characteristic array and each second characteristic array, the computer system first performs the combination operation.

[0087] For the first feature array, it is assumed that multiple first feature arrays are obtained by feature representation of the observation results corresponding to the computer room facility parameters and the temperature history control data. For example, a first feature array A1 is obtained for the server power in the computer room facility parameters. It is assumed that after the value range of the server power is mapped to the feature value, A1=[1, 3, 2,4] (each number here represents the feature value corresponding to different server powers); the first feature array A2=[5, 3, 4, 2] is obtained for the number of servers; the first feature array A3=[3, 2, 1, 2] is obtained for the air conditioning setting temperature in the temperature history control data. The computer system combines these first feature arrays into a fourth fused feature array. The combination method can be a simple sequential splicing, then the fourth fused feature array F4=[1, 3, 2, 4, 5, 3, 4, 2, 3, 2, 1,2]. For the second feature array, it is also assumed that multiple second feature arrays are obtained by feature representation of the time series observation results corresponding to different temperature history control time series data. Taking the control time series data of the temperature control device triggered by refrigeration as an example, the second feature array B1 obtained by feature representation is assumed to be an array with a larger dimension after one-hot encoding. In order to simplify the example, it is assumed that B1=[0, 1, 0, 0, 1]; for the second feature array B2 corresponding to the control time series data of the temperature control device triggered by fan speed increase=[1, 0, 1, 0, 0]. The computer system combines these second feature arrays into the fifth fused feature array, and obtains F5=[0, 1, 0, 0, 1, 1, 0, 1, 0, 0] after splicing in order.

[0088] The purpose of this combination operation is to integrate the scattered feature arrays related to different aspects, in preparation for further mining the relationship between features. By combining the first feature arrays into the fourth fused feature array, the computer system integrates the feature information related to the room facility parameters and the temperature history control data into one data structure. Similarly, the second feature arrays are combined into the fifth fused feature array to integrate the feature information related to the temperature history control time series data.

[0089] In step S220, the computer system uses multiple indicator combination models to process the fourth fusion feature array F4=[1, 3,2, 4, 5, 3, 4, 2, 3, 2, 1, 2]. The indicator combination model is a model specifically used to analyze the combination relationship between different indicators in the feature array. Assume that there are three indicator combination models, namely models M1, M2 and M3.

[0090] Model M1, for example, adopts an indicator combination model based on linear regression. It assumes that there is a certain linear relationship between the elements in the fourth fusion feature array. For example, it establishes a linear relationship model between the server power characteristic values ​​(such as elements in A1) and the server quantity characteristic values ​​(such as elements in A2) and the air conditioning set temperature characteristic values ​​(such as elements in A3) in the array. The computer system inputs F4 into model M1, and model M1 calculates the elements in F4 according to pre-set linear regression coefficients (assuming that these coefficients are set through previous training or based on domain knowledge). It is assumed that the calculation process of model M1 is to combine and calculate the first four elements in F4 (corresponding to the server power characteristic values) with the fifth to eighth elements (corresponding to the server quantity characteristic values) and the ninth to twelfth elements (corresponding to the air conditioning set temperature characteristic values) according to a linear relationship. After calculation, the first array output by model M1 is expressed as R1=[0.5, 0.3, 0.2]. Each element here is calculated by model M1 based on its internal linear relationship, and represents the results of different indicator combination relationships. For example, the first element 0.5 may represent a comprehensive indicator relationship of the impact of server power and server quantity on temperature control.

[0091] Model M2, for example, adopts an indicator combination model based on principal component analysis (PCA). Its purpose is to find the main component (principal component) in the fourth fusion feature array that best represents the data change, so as to determine the indicator combination relationship. The computer system inputs F4 into model M2. Model M2 first standardizes F4, then calculates the covariance matrix, and then finds the eigenvalues ​​and eigenvectors. Assume that after these calculations, model M2 determines an indicator combination relationship based on the main eigenvectors, and the output first array is represented as R2=[0.4, 0.4, 0.2], where the elements represent the weights or scores of different indicator combination relationships obtained based on principal component analysis.

[0092] Model M3, for example, adopts an indicator combination model based on a decision tree. It uses the elements in F4 as input features and constructs a decision tree to analyze the combination relationship between different indicators. For example, the root node of the decision tree may be a threshold judgment of the server power characteristic value, the branch node may be a different range judgment of the server quantity characteristic value, and the leaf node is a certain result related to the air conditioner set temperature characteristic value. The computer system inputs F4 into model M3, and model M3 calculates according to the structure of the decision tree. The first array of the final output is represented as R3=[0.3, 0.5, 0.2], where the elements represent the results of the combination relationship of different indicators obtained by the decision tree model analysis.

[0093] These different indicator combination models analyze the indicator combination relationship in the fourth fusion feature array from different perspectives, and the output first array representation provides different bases for further fusion of feature information.

[0094] In step S230, the computer system loads the fifth fused feature array F5=[0, 1, 0, 0, 1, 1, 0, 1, 0,0] into the sequence processing recursive model, where the gated recurrent unit (GRU) is taken as an example. GRU is a special recurrent neural network (RNN) that can process sequence data and capture long-term dependencies in the data. When the computer system loads F5 into the GRU, the input layer of the GRU receives the elements in F5. The GRU has structures such as update gates and reset gates to control the flow and update of information. Assume that the hidden layer of the GRU has 3 neurons. At each time step (for each element in F5), the GRU updates the hidden state according to the current input element, the hidden state of the previous time step, and the calculation results of the update gate and reset gate. For example, for the first element 0 in F5, the GRU calculates the values ​​of the update gate and reset gate according to its initial hidden state (assuming that it is initialized as a full 0 vector) and the internal weight matrix and bias vector, and then updates the hidden state according to these values. As the elements in F5 are processed one by one, the GRU continuously updates the hidden state.

[0095] After processing the entire F5, the output layer of GRU will output a result, which is the second array representation. Assume that this second array is represented as S=[0.2, 0.3, 0.1]. Each element here is the feature information related to the sequence mined from the temperature history control time series data by GRU based on the processing of F5. For example, the first element 0.2 may represent the influence weight of a certain feature in the initial stage on the current temperature control in the entire temperature control time series; the second element 0.3 may represent the relevant influence weight in the middle stage; the third element 0.1 may represent the influence weight in the later stage.

[0096] In step S240, the computer system combines the first array representations of the outputs of the various indicator combination models obtained previously (such as R1=[0.5, 0.3, 0.2], R2=[0.4, 0.4, 0.2], R3=[0.3, 0.5, 0.2]) and the second array representation S=[0.2, 0.3, 0.1] output by the sequence processing recursive model into a first fused feature array.

[0097] The combination method can be direct concatenation or weighted combination. If the direct concatenation method is adopted, the first fusion feature array F1=[0.5, 0.3, 0.2, 0.4, 0.4, 0.2, 0.3, 0.5, 0.2, 0.2, 0.3,0.1]. If the weighted combination method is adopted, it is assumed that the weight represented by the first array is 0.6 and the weight represented by the second array is 0.4. For the first element after combination, the calculation method is: (0.5×0.6+0.2×0.4)=0.3+0.08=0.38; for the second element: (0.3×0.6+0.3×0.4)=0.18+0.12=0.3; for the third element: (0.2×0.6+0.1×0.4)= 0.12+0.04=0.16, and so on, to obtain the first fused feature array after weighted combination.

[0098] Through this combination, the first fusion feature array integrates various feature relationship information mined from the equipment room facility parameters, temperature history control data, and temperature history control time series data. This fused feature array can more comprehensively reflect the various factors related to the temperature control of the equipment room and their interrelationships, providing richer and more accurate input information for subsequent steps (such as loading into the first task model for feature extraction, etc.), which helps to improve the accuracy and reliability of the entire equipment room temperature prediction control.

[0099] As an implementation mode, step S400, determining the integrity of the control evaluation factor data in each pre-set evaluation factor set, and generating a primary data integrity array representation of the control evaluation factor data according to the integrity of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, includes:

[0100] Step S410: for an evaluation element set, determine the number of blank evaluation elements corresponding to the blank observation results in the evaluation element set, subtract the number of blank evaluation elements from the total number of evaluation elements in the evaluation element set to obtain the total number of complete evaluation elements, divide the total number of complete evaluation elements by the total number of evaluation elements to obtain the completeness of the evaluation element set;

[0101] Step S420: For each evaluation factor set, the corresponding completeness is calculated according to the above process;

[0102] Step S430: for each evaluation factor set, multiply its corresponding integrity by its set feature array to obtain the adjustment feature array corresponding to each evaluation factor set, and combine the adjustment feature arrays corresponding to each evaluation factor set to obtain a primary data integrity array representation.

[0103] In step S410, the evaluation factor set is a set of control evaluation factor data classified and combined according to specific rules or logic. For example, suppose there is an evaluation factor set about the parameters of the computer room facilities, and the evaluation factors included in this set are the number of servers, server power, computer room space size, insulation material performance, and the number of ventilation equipment.

[0104] When the computer system checked the control evaluation factor data for the room, it found that there were no observations for the insulation material performance factor, which means that there was missing data. At this time, the number of blank evaluation factors corresponding to the blank observation results is 1. The total number of evaluation factors in this evaluation factor set is 5. Next, the computer system calculates the total number of complete evaluation factors, subtracting the number of blank evaluation factors from the total number of evaluation factors, that is, 5-1=4. Finally, the total number of complete evaluation factors is divided by the total number of evaluation factors to obtain the completeness of the evaluation factor set, which is calculated as 4÷5=0.8.

[0105] Let's look at another evaluation factor set, for example, the evaluation factor set for temperature history control data, which includes the four elements of the historical set temperature of air conditioner 1, the cooling power of air conditioner 1, the historical speed of fan 1, and the historical set temperature of air conditioner 2. If there is no record of the historical set temperature of air conditioner 2 in the control evaluation factor data, then the number of blank evaluation factors is 1, and the total number of evaluation factors is 4. The total number of complete evaluation factors is 4-1=3, and the completeness of the evaluation factor set is 3÷4=0.75.

[0106] For the temperature history control time series data evaluation element set, for example, in the cooling trigger time series data dimension, it is assumed that the three elements are the equipment code of the last three cooling triggers, the time interval of the cooling triggers, and the ambient temperature when the cooling triggers. If there is no observation result of the ambient temperature when the cooling is triggered, the number of blank evaluation elements is 1, the total number of evaluation elements is 3, and the total number of complete evaluation elements is 3-1=2, and its completeness is 2÷3≈0.67.

[0107] This method of calculating the completeness of the evaluation factor set can quantify the data completeness of each evaluation factor set, providing an important basis for subsequent processing. Different evaluation factor sets may have different data completeness due to data sources, collection equipment or other factors. This method can clarify the extent of such differences.

[0108] In step S420, the computer system performs the operation of calculating the completeness in S410 for all evaluation factor sets. Assume that in addition to the three evaluation factor sets of the equipment room facility parameters, temperature history control data, and temperature history control time series data mentioned above, there are other evaluation factor sets, such as the evaluation factor set related to the power supply of the equipment room (including total power input, the proportion of power allocated to each device, power supply stability, etc.) and the evaluation factor set related to the external environment of the equipment room (including the orientation of the building where the equipment room is located, the surrounding environment temperature, the surrounding environment humidity, etc.).

[0109] For the evaluation factor set related to the power supply of the computer room, it is assumed that it contains 3 factors. When checking the control evaluation factor data, it is found that there is no observation result for the proportion of power distributed to each device. The number of blank evaluation factors is 1, and the total number of evaluation factors is 3. The total number of complete evaluation factors is calculated to be 3-1=2, and the completeness of the evaluation factor set is 2÷3≈0.67. For the evaluation factor set related to the external environment of the computer room, it is assumed that it contains 4 factors. If there is no observation result for the surrounding environmental humidity, the number of blank evaluation factors is 1, and the total number of evaluation factors is 4. The total number of complete evaluation factors is calculated to be 4-1=3, and its completeness is 3÷4=0.75.

[0110] By performing such completeness calculations on each evaluation factor set, the computer system can fully grasp the data integrity of each evaluation factor set, which is very important for comprehensively considering the quality and reliability of the data related to the temperature control of the entire computer room. The completeness of different evaluation factor sets reflects the integrity of data in various aspects. This information will be used in subsequent steps to generate a primary data integrity array representation, which in turn affects the prediction process of the temperature control parameters of the entire computer room.

[0111] In step S430, each evaluation factor set has a pre-set feature array, which is set according to the characteristics of the evaluation factor set and the subsequent processing requirements. For example, for the computer room facility parameter evaluation factor set, the set feature array may be [2, 3, 1, 4, 2]. The completeness of the evaluation factor set is calculated to be 0.8. The computer system multiplies the completeness with each element in the feature array to obtain the adjusted feature array [2×0.8, 3×0.8, 1×0.8,4×0.8, 2×0.8]=[1.6, 2.4, 0.8, 3.2, 1.6]. For the evaluation factor set of temperature history control data, assuming that its feature array is [3, 1, 2, 1], the completeness is 0.75, and after multiplication, the adjustment feature array is [3×0.75, 1×0.75, 2×0.75, 1×0.75]=[2.25, 0.75, 1.5, 0.75]. For the evaluation factor set of temperature history control timing data, assuming that its feature array is [1, 2, 3], the completeness is 0.67, and after multiplication, the adjustment feature array is [1×0.67, 2×0.67, 3×0.67]=[0.67, 1.34, 2.01]. For the evaluation factor set related to the power supply of the computer room, assuming that its feature array is [1, 2, 1] and the completeness is 0.67, the adjustment feature array is [1×0.67, 2×0.67, 1×0.67]=[0.67, 1.34, 0.67]. For the evaluation factor set related to the external environment of the computer room, assuming that its feature array is [2, 1, 3, 1] and the completeness is 0.75, the adjustment feature array is [2×0.75, 1×0.75, 3×0.75, 1×0.75]=[1.5, 0.75, 2.25, 0.75]. The computer system combines the adjustment feature arrays corresponding to each evaluation factor set to obtain the primary data integrity array representation. Assuming that the evaluation factor sets are combined in the order mentioned above, the resulting primary data completeness array is expressed as [1.6, 2.4, 0.8, 3.2, 1.6, 2.25, 0.75, 1.5,0.75, 0.67, 1.34, 0.67, 1.5, 0.75, 2.25, 0.75].

[0112] This primary data integrity array representation integrates the data integrity information of each evaluation element set and the characteristic array information related to the characteristics of each evaluation element set. It can identify the blank content of the observation results corresponding to the reference indicators in the control evaluation element data. For example, if the value of an element in the primary data integrity array representation is low, it may mean that there is a lot of missing or incomplete data in the corresponding evaluation element set. This can serve as important reference information in the subsequent model processing process, allowing the model to be adjusted accordingly according to the integrity of the data, thereby improving the accuracy and reliability of the entire room temperature prediction control.

[0113] As an implementation mode, step S512 generates a secondary data integrity array representation of the control evaluation factor data based on the integrity of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, including:

[0114] Step S5121: for each pair of evaluation factor sets, according to the completeness of the control evaluation factor data in each evaluation factor set in the pair of evaluation factor sets, determine the completeness of the indicator combination of the control evaluation factor data in the pair of evaluation factor sets, and according to the feature array set for each evaluation factor set in the pair of evaluation factor sets, determine the feature array of the pair of evaluation factor sets;

[0115] Step S5122: Determine the secondary data integrity array representation of each pair of evaluation element sets based on the indicator combination completeness and feature array of each pair of evaluation element sets; combine the secondary data integrity array representation of each pair of evaluation element sets to obtain the secondary data integrity array representation of the control evaluation element data.

[0116] In step S5121, it is assumed that there are three evaluation element sets: the equipment room facility parameter evaluation element set (referred to as set A), the temperature history control data evaluation element set (referred to as set B), and the temperature history control time series data evaluation element set (referred to as set C). For set A, the previously calculated completeness is 0.8, and its set feature array is [2, 3, 1, 4, 2]; for set B, the completeness is 0.75, and the feature array is [3, 1, 2, 1]; for set C, the completeness is 0.67, and the feature array is [1, 2, 3].

[0117] The computer system begins to process each pair of evaluation factor sets. Let's look at the pair of set A and set B first. According to the rules, to determine the completeness of the indicator combination of this pair of evaluation factor sets, it is necessary to calculate the product of the completeness of the two evaluation factor sets. That is, the completeness of the indicator combination of set A and set B is 0.8×0.75=0.6. Next, let's look at the pair of set A and set C, whose indicator combination completeness is 0.8×0.67=0.536. Finally, let's look at the pair of set B and set C, whose indicator combination completeness is 0.75×0.67=0.5025.

[0118] This method of calculating the completeness of the indicator combination can reflect the comprehensive situation of the data completeness between the two evaluation factor sets. For example, for set A and set B, if the data completeness of one of the sets is low, then their indicator combination completeness will also be low, which means that when considering the indicator combination relationship between the two sets, the reliability of the data is relatively low.

[0119] For the pair of sets A and B, to determine their feature arrays, we need to calculate the dot product of the feature arrays of the two evaluation feature sets. The feature array of set A is [2, 3, 1, 4, 2], and the feature array of set B is [3, 1, 2,1]. The calculation is as follows: calculation of the first element: 2×3=6; calculation of the second element: 3×1=3; calculation of the third element: 1×2=2; calculation of the fourth element: 4×1=4; fifth element (since the feature array of set B has only four elements, it is calculated as 0 here): 2×0=0. Adding these results together, we get the feature array of the pair of evaluation feature sets as [6+3+2+4+0]=15.

[0120] For the pair of sets A and C, the feature array of set A is [2, 3, 1, 4, 2], and the feature array of set C is [1, 2, 3]. Calculate the dot product: the first element: 2×1=2; the second element: 3×2=6; the third element: 1×3=3; the fourth element (since the feature array of set C has only three elements, it is calculated as 0 here): 4×0=0; the fifth element: 2×0=0; add them together to get the feature array of this pair of evaluation feature sets: [2+6+3+0+0]=11.

[0121] For the pair of set B and set C, the feature array of set B is [3, 1, 2, 1], and the feature array of set C is [1, 2, 3]. Calculate the dot product: the first element: 3×1=3; the second element: 1×2=2; the third element: 2×3=6; the fourth element (since the feature array of set C has only three elements, it is calculated as 0 here): 1×0=0. Adding them together, the feature array of this pair of evaluation feature sets is [3+2+6+0]=11.

[0122] Determining the feature array of each pair of evaluation factor sets helps to consider the specific relationship between the indicator combinations of different evaluation factor sets in the subsequent steps. Different feature array values ​​reflect the intrinsic connection and weight relationship between different evaluation factor sets.

[0123] In step S5122, the secondary data integrity array representation of each pair of evaluation factor sets is determined. For the pair of set A and set B, the computer system multiplies the previously calculated indicator combination completeness 0.6 and the feature array 15 to obtain a part of the secondary data integrity array representation of this pair, that is, 0.6×15=9. For the pair of set A and set C, the indicator combination completeness is 0.536, the feature array is 11, and the multiplication is 0.536×11=5.896. For the pair of set B and set C, the indicator combination completeness is 0.5025, the feature array is 11, and the multiplication is 0.5025×11=5.5275.

[0124] These results constitute the secondary data integrity array representation of each pair of evaluation factor sets. These representations reflect the comprehensive situation of data integrity and the relationship between specific indicators when considering the relationship between different evaluation factor sets. For example, for the result 9 of set A and set B, if this value is relatively large, it means that between the two evaluation factor sets, the data integrity and the indicator combination relationship between them have a relatively large influence in the whole; if the value is small, it means that there may be more uncertainty in this regard or that the data missing has a greater impact on the relationship between the two.

[0125] The computer system combines the secondary data integrity array representations of each pair of evaluation factor sets to obtain the secondary data integrity array representation of the control evaluation factor data. Assuming that the combination is in the order mentioned above, the secondary data integrity array representation obtained is [9, 5.896, 5.5275].

[0126] This secondary data integrity array representation can identify the blank content of the indicator combination between the reference indicators in the control evaluation factor data. For example, if one of the values ​​is low, it may mean that there are more blank indicator combinations between the corresponding two evaluation factor sets, that is, the data of some key indicator combinations are missing or incomplete. This will affect the subsequent accurate prediction of the temperature control parameters of the computer room, because in the model of computer room temperature control, the indicator combination relationship between different evaluation factor sets is very important for accurately predicting the temperature control parameters. If there is incomplete data in some indicator combination relationships, the model may not be able to accurately capture the impact of these relationships on temperature control, resulting in deviations in the prediction results. Through this secondary data integrity array representation, the computer system can better consider these factors in subsequent processing, such as in operations such as determining the influence coefficient of the second task model, and make reasonable adjustments based on the integrity of the data to improve the accuracy of the prediction of the temperature control parameters of the entire computer room.

[0127] As an implementation mode, step S5121, according to the completeness of each evaluation element set of the control evaluation element data in the pair of evaluation element sets, determines the completeness of the index combination of the control evaluation element data in the pair of evaluation element sets, and according to the feature array set for each evaluation element set in the pair of evaluation element sets, determines the feature array of the pair of evaluation element sets, including:

[0128] Step S51211: determining the product of the completeness of each evaluation element set in the pair of evaluation element sets as the indicator combination completeness of the control evaluation element data in the pair of evaluation element sets;

[0129] Step S51212: Determine the dot product of the feature arrays set for each evaluation factor set in the pair of evaluation factor sets as the feature array of the pair of evaluation factor sets.

[0130] Step S5122, determining the secondary data integrity array representation of each pair of evaluation element sets according to the indicator combination integrity and feature array of each pair of evaluation element sets, including:

[0131] Step S51221: Multiply the indicator combination integrity and feature array of each pair of evaluation factor sets to obtain the secondary data integrity array representation of each pair of evaluation factor sets.

[0132] The principles of steps S51211 to S51212 have been explained in the above example and will not be repeated here.

[0133] In step S51221, the computer system determines the secondary data integrity array representation based on the indicator combination completeness and feature array of each pair of evaluation element sets calculated previously. For the pair of set A and set B, the indicator combination completeness obtained previously is 0.6, and the feature array is 15. The computer system multiplies these two values: 0.6×15=9. This 9 is part of the secondary data integrity array representation of the pair of evaluation element sets A and set B. For the pair of set A and set C, the indicator combination completeness is 0.536, and the feature array is 11. Multiply to get: 0.536×11=5.896. This is the secondary data integrity array representation part of the pair of set A and set C. For the pair of set B and set C, the indicator combination completeness is 0.5025, and the feature array is 11. Multiply to get: 0.5025×11=5.5275. This is the secondary data integrity array representation part of the pair of set B and set C. These secondary data integrity array representations reflect the comprehensive situation of the data integrity and the combination relationship of specific indicators when considering the relationship between different evaluation factor sets. For example, the secondary data integrity array representation corresponding to set A and set B is 9. If this value is relatively large, it means that between the two evaluation factor sets, the data integrity and the combination relationship of indicators between them have a relatively large influence in the whole; if the value is small, it means that there may be more uncertainty in this regard or that the data missing has a greater impact on the relationship between the two. These secondary data integrity array representations will be further combined to obtain the secondary data integrity array representation of the control evaluation factor data, thereby providing a basis for subsequent processing (such as determining the influence coefficient of the second task model, etc.), which helps to improve the accuracy of the prediction of the temperature control parameters of the entire computer room.

[0134] As an implementation mode, the method is implemented by a temperature control parameter prediction neural network. Before obtaining the control evaluation factor data of the computer room to be controlled, the method also includes a training process of the temperature control parameter prediction neural network, specifically including:

[0135] Step S10: obtaining a calibration sample library and a result verification sample library, both of which are training databases including sample data of multiple computer rooms and prior information corresponding to the sample data of each computer room, the computer rooms corresponding to the sample data of the computer rooms in the calibration sample library and the result verification sample library are computer rooms in different areas, the computer room sample data include observation results corresponding to the computer room and multiple computer room facility parameters, observation results corresponding to the computer room and multiple temperature history control data, and observation results corresponding to the computer room and multiple temperature history control time series data, the temperature history control data include temperature control parameters of the computer room on multiple temperature control devices, the temperature history control time series data represent the temperature control device control time series data obtained based on the control of the target number of times of the computer room on multiple temperature control devices before the current moment, and the prior information is the temperature control parameters of the computer room for the target temperature control device in a preset period;

[0136] Step S20: Debugging and obtaining a temperature control parameter prediction neural network based on a calibration sample library;

[0137] Step S30: verifying the temperature control parameter prediction neural network based on the result verification sample library, wherein obtaining the control evaluation factor data of the computer room to be controlled is performed when the temperature control parameter prediction neural network is based on the verification.

[0138] In step S10, the computer system obtains a calibration sample library and a result verification sample library. Both sample libraries are important data sources for training and verifying the temperature control parameter prediction neural network.

[0139] The calibration sample library contains sample data from multiple computer rooms and prior information corresponding to the sample data from each computer room. For example, suppose the calibration sample library contains data from five different computer rooms (computer room 1, computer room 2, computer room 3, computer room 4, computer room 5). For computer room 1, the observation results of the computer room facility parameters may include the number of servers is 50, the total server power is 10,000 watts, the computer room space is 100 square meters, etc.; the temperature history control data may be the historical set temperatures of air conditioner 1 at the past 10 time points are [20, 21, 22, 20, 19, 20, 21, 20, 19, 20] degrees Celsius, the cooling power is [50%, 40%, 45%, 50%, 55%, 50%, 45%, 50%, 55%, 50%], etc., and the historical fan speed at the corresponding time points is [1000, 1200, 1100, 1000, 900, 1000, 1100, 1000, 900, 1000] rpm; Taking the cooling trigger as an example, the temperature history control time series data assumes that the target number is 5 times, and the last 5 cooling trigger device codes are [101, 102, 101, 0, 102] (if less than 5 times, use 0 to supplement). The corresponding prior information may be that in the next preset period (such as the next hour), the temperature control parameters of air conditioner 1 are set to 20 degrees Celsius and the cooling power is 50%.

[0140] The result verification sample library also has a similar structure, but the room sample data corresponds to a room different from the room in the calibration sample library to ensure the independence of verification. For example, the result verification sample library contains data from three other rooms (room 6, room 7, and room 8), and the composition of the room sample data is the same as that of the calibration sample library, including room facility parameters, temperature history control data, and temperature history control time series data, and also has their own corresponding prior information.

[0141] The computer room facility parameters in the computer room sample data reflect the basic hardware conditions of the computer room, and these parameters will have an impact on the temperature of the computer room. For example, a computer room with a large number of servers and high power may generate more heat and require more cooling to maintain a stable temperature. The temperature history control data records the operation of the temperature control equipment in the past. By analyzing these historical data, the rules of temperature control can be found. The temperature history control time series data provides information such as the order of operation of the temperature control equipment from the perspective of time series, which helps to capture the dynamic characteristics of temperature control. Prior information is the known target temperature control parameters within a specific preset period. It is the target value for model training and verification. The computer system tries to make the model output close to the prior information by learning and adjusting the data in the sample library, and then uses the result to verify the sample library to test the accuracy of the model.

[0142] In step S20, the computer system uses the data in the calibration sample library to debug the temperature control parameter prediction neural network. It is assumed that the initial temperature control parameter prediction neural network has a specific structure, for example, including an input layer, a hidden layer, and an output layer. The number of neurons in the input layer needs to match the characteristic dimension of the computer room sample data. Taking the computer room sample data mentioned above as an example, the computer room facility parameters, temperature history control data, and temperature history control time series data of computer room 1 are integrated and preprocessed to form a feature vector as input. For example, the integrated feature vector may be a vector with a dimension of 100 (this is the result of encoding, standardization, etc. of various data). The computer system loads this feature vector into the input layer of the initial temperature control parameter prediction neural network. When the data enters the input layer, it is processed by the hidden layer. The neurons in the hidden layer are connected by weights, and the activation function is used to introduce nonlinear characteristics. For example, it is assumed that there are 50 neurons in the hidden layer and the activation function is ReLU (linear rectification function). Each neuron performs weighted summation on the input feature vector according to its weight and performs nonlinear transformation through the ReLU function. This process is passed layer by layer between multiple neurons in the hidden layer, continuously extracting and abstracting features in the data. In the output layer, the neural network outputs results related to the prediction of temperature control parameters. For example, the output may be a vector containing the predicted temperature setting value and cooling power. Assume that for the input data of room 1, the prediction result output by the initial temperature control parameter prediction neural network is that the temperature setting value is 21 degrees Celsius and the cooling power is 45%, while the corresponding prior information is that the temperature setting value is 20 degrees Celsius and the cooling power is 50%.

[0143] The computer system adjusts the parameters of the neural network (such as weights and biases) based on the difference between the output and the prior information. Various optimization algorithms can be used, such as the stochastic gradient descent (SGD) algorithm. Assuming that the mean square error (MSE) is used as the loss function, calculate the MSE between the predicted result and the prior information. For the temperature setting value, the predicted value is 21 degrees Celsius and the prior value is 20 degrees Celsius, the error is (21-20) 2 =1; for cooling power, the predicted value is 45%, the priori value is 50%, and the error is (0.45-0.5) 2 = 0.0025. The overall MSE is the average of these two errors (this is just a simple example, the actual calculation may involve more data and more complex calculations).

[0144] According to the calculated loss value, the SGD algorithm is used to update the weights and biases of the neural network. For example, if the current value of a weight is 0.5, according to the update rule of the SGD algorithm, the new weight value is calculated to be 0.5-learning rate × the derivative of the loss function with respect to the weight (the learning rate here is a pre-set small positive number, such as 0.01). By continuously inputting the computer room sample data in the calibration sample library into the neural network, calculating the loss function and updating the parameters, the prediction results of the neural network are gradually made closer to the prior information, thereby completing the debugging of the temperature control parameter prediction neural network.

[0145] In step S30, after completing the debugging based on the calibration sample library, the computer system uses the result verification sample library to verify the debugged temperature control parameter prediction neural network.

[0146] For the computer rooms in the result verification sample library (such as computer room 6), the computer system processes its computer room sample data in the same way as during calibration and loads it into the temperature control parameter prediction neural network. For example, the integrated feature vector of computer room 6 (assuming the dimension is 100) is loaded into the input layer of the neural network, and the prediction result is obtained at the output layer after processing by the hidden layer. Assume that for computer room 6, the predicted temperature setting value is 19 degrees Celsius and the cooling power is 48%, while the corresponding prior information is that the temperature setting value is 20 degrees Celsius and the cooling power is 50%.

[0147] Also use some error metric (such as mean square error MSE) to evaluate the difference between the prediction result and the prior information. For the temperature setting value, the error is (19-20) 2 =1; for cooling power, the error is (0.48-0.5) 2 =0.0004. The total MSE is the average of these two errors. By calculating the MSE between the prediction results and the prior information of all the rooms in the result verification sample library, the average error of the entire result verification sample library can be obtained. If this average error is within an acceptable range, it means that the temperature control parameter prediction neural network has good accuracy; if the average error is too large, the neural network may need to be further adjusted (such as returning to the calibration sample library for more debugging or adjusting the structure of the neural network, etc.).

[0148] During this verification process, obtaining the control evaluation factor data of the computer room to be controlled will also be performed. This is because in actual applications, the data structure and format of the computer room to be controlled are similar to the computer room sample data in the result verification sample library. By obtaining the control evaluation factor data of the computer room to be controlled during the verification process, it can be ensured that when the neural network is applied to the actual computer room temperature control, the model has been fully verified and can correctly handle similar data structures. For example, if a parameter in the control evaluation factor data of the computer room to be controlled (such as the representation of server power) is different from the data in the result verification sample library, it can be discovered during the verification process and corresponding adjustments can be made (such as adjustments to the data preprocessing method), thereby ensuring the effectiveness of the temperature control parameter prediction neural network in actual applications.

[0149] As an implementation manner, in step S30, before verifying the temperature control parameter prediction neural network based on the result verification sample library, the method may further include:

[0150] Step S301: obtaining a process verification sample library including computer room sample data of multiple computer rooms and prior information corresponding to each computer room sample data;

[0151] Step S302: cyclically updating the temperature control parameter prediction neural network based on the calibration sample library, and after each update, verifying the temperature control parameter prediction neural network in the calibration process according to the process verification sample library to obtain the neural network debugging accuracy;

[0152] Step S303: after each update, determine whether the neural network debugging accuracy obtained after multiple consecutive updates has converged, wherein the temperature control parameter prediction neural network is verified based on the result verification sample library when it is determined that the neural network debugging accuracy obtained after multiple consecutive updates has converged.

[0153] In step S301, the computer system obtains a process verification sample library, which also contains relevant data of multiple computer rooms. Its structure and purpose are related to the previously mentioned adjustment sample library and result verification sample library.

[0154] Assume that the process validation sample library contains data from 10 different computer rooms (computer room A-computer room J). For computer room A, the computer room facility parameters in its computer room sample data may include the number of servers is 30, the average server power is 800 watts, the thermal insulation performance index of the computer room is 0.7 (assuming this is a quantitative value to measure the thermal insulation effect), etc. In terms of temperature history control data, the historical set temperatures of the air conditioner at the past 8 time points are [22, 23, 21, 22, 20, 21, 23, 22] degrees Celsius, the cooling power is [40%, 45%, 40%, 45%, 50%, 45%, 40%, 45%], and the historical speed of the fan at the corresponding time points is [900, 1000, 900, 1000, 1100, 1000, 900, 1000] rpm. Taking the fan speed trigger as an example, the temperature history control time series data assumes that the target number is 5 times, and the device codes for the last 5 fan speed triggers are [201, 202, 201, 0, 202]. The corresponding prior information may be that in the next preset period (for example, the next 45 minutes), the temperature control parameters of the air conditioner are set to 22 degrees Celsius, the cooling power is 45%, and the fan speed is 1000 rpm.

[0155] For other computer rooms (such as computer room B-computer room J), there are also computer room sample data with similar structures and corresponding prior information. These data cover the different states and historical control situations of the computer rooms, providing rich materials for the subsequent verification process.

[0156] The various parameters and data in the computer room sample data reflect the actual operating status and historical control status of the computer room. The parameters of the computer room facilities affect the heat generation and heat dissipation requirements of the computer room. For example, the number and power of servers will generate heat, and the insulation performance will affect the rate of heat dissipation. The temperature history control data and temperature history control time series data record the operation history of the temperature control equipment, which helps to analyze the control mode and law of the temperature control equipment. Prior information is the known target temperature control parameters within a specific time period. It is used as a standard to compare with the model prediction results to judge the accuracy of the model.

[0157] In step S302, the computer system uses the calibration sample library to train the temperature control parameter prediction neural network. Assume that the initial temperature control parameter prediction neural network has a certain structure, for example, the input layer has 100 neurons (corresponding to the integrated room sample data features), the hidden layer has 50 neurons, the activation function is ReLU, and the output layer outputs the temperature control parameter prediction results (such as temperature setting value and cooling power, etc.).

[0158] During the training process, the computer system inputs the sample data of the computer room in the calibration sample library into the neural network one by one. Taking the data of computer room 1 as an example, its integrated feature vector is input into the input layer of the neural network. The data is processed through the hidden layer in the neural network, and the neurons are weighted and summed according to the pre-set weights, and nonlinear transformation is performed through the ReLU function. After obtaining the prediction result in the output layer, the computer system calculates the error based on the prediction result and the corresponding prior information. For example, the mean square error (MSE) is used as the error metric. If the predicted temperature setting value is 21 degrees Celsius and the prior value is 20 degrees Celsius, the error is (21-20) 2 =1; for cooling power, assuming the predicted value is 45%, the prior value is 50%, and the error is (0.45-0.5) 2 = 0.0025. The overall MSE is the average of these two errors (this is just a simple example, the actual calculation may be more complicated).

[0159] The computer system updates the weights and biases of the neural network based on the calculated error using an optimization algorithm (such as the stochastic gradient descent algorithm). Assuming the initial weights are [w1, w2, …, wn], according to the update rule of the stochastic gradient descent algorithm, the new weights are [w1-learning rate × MSE derivative of w1, w2-learning rate × MSE derivative of w2, …, wn-learning rate × MSE derivative of wn] (where the learning rate is a pre-set small positive number, such as 0.01).

[0160] The computer system will perform multiple such update cycles, continuously adjusting the parameters of the neural network to make the neural network's predictions closer to the prior information.

[0161] After each update, the computer system uses the process verification sample library to verify the neural network during the calibration process. For room A in the process verification sample library, the computer system processes its room sample data in the same way as during calibration and inputs it into the updated neural network. For example, the integrated feature vector of room A is input into the neural network to obtain the predicted result (assuming that the predicted temperature setting value is 22.5 degrees Celsius and the cooling power is 43%), which is compared with the corresponding prior information (the temperature setting value is 22 degrees Celsius and the cooling power is 45%). The mean square error (MSE) is also used to calculate the error. For the temperature setting value, the error is (22.5-22) 2 =0.25; for cooling power, the error is (0.43-0.45) 2 = 0.0004. The overall MSE is the average of these two errors.

[0162] By performing such predictions and error calculations for all the computer rooms (computer rooms A to J) in the process validation sample library, the computer system can obtain an average error value. The inverse of this average error value or other error-related conversions (such as 1-average error value, if the average error value is between 0 and 1) can be used as a measure of the accuracy of neural network debugging. For example, if the average error value is 0.1, then the neural network debugging accuracy can be expressed as 1-0.1=0.9, indicating that the neural network has 90% accuracy in this verification process.

[0163] In step S303, after each update of the neural network, the computer system will record the neural network debugging accuracy obtained according to the process verification sample library. Assume that in the process of 10 consecutive updates, the neural network debugging accuracy obtained is [0.8, 0.82, 0.83, 0.835, 0.84, 0.842, 0.843, 0.844, 0.844, 0.844]. To determine whether it has converged, a variety of methods can be used. One feasible method is to observe the trend of accuracy changes. If the change in accuracy is less than a preset threshold (for example, 0.001) after multiple consecutive updates, it can be considered that the neural network debugging accuracy has converged. In this example, from the 8th update to the 10th update, the change in accuracy is very small (all 0.844). If the preset threshold is 0.001, it can be considered that the accuracy has converged.

[0164] Another method is based on statistical tests, such as using a t-test to determine whether the accuracies after multiple consecutive updates come from the same distribution. If the results of the t-test show that there is no statistically significant difference between these accuracies, it can also be considered that the accuracies have converged.

[0165] When it is determined that the neural network debugging accuracy has converged, the computer system verifies the temperature control parameter prediction neural network based on the result verification sample library. This is because it is meaningful to use the result verification sample library for verification only when the neural network reaches a certain stable state (i.e., the accuracy converges) during the adjustment process. If the result verification is performed before the neural network converges, inaccurate verification results may be obtained, because the neural network may still be in the process of continuous adjustment and its performance is not stable. For example, if the result verification sample library is used for verification before the accuracy has converged, it may be due to problems such as overfitting or underfitting of the neural network that the performance of the neural network is very good or very poor, while in fact the neural network has not yet reached its optimal state. By first using the process verification sample library to continuously verify and determine the accuracy convergence during the adjustment process, and then using the result verification sample library for final verification, the performance of the temperature control parameter prediction neural network can be more accurately evaluated to ensure that it can accurately predict the temperature control parameters of the computer room in actual applications.

[0166] As an implementation mode, step S20, debugging and obtaining a temperature control parameter prediction neural network based on a calibration sample library, includes:

[0167] Step S21: Loading the computer room sample data in the calibration sample library into the initial temperature control parameter prediction neural network to obtain the training confidence, training expectation and training variance deviation output by the initial temperature control parameter prediction neural network;

[0168] Step S22: determining a first error of the information gain error function based on the training confidence and the prior information corresponding to the computer room sample data;

[0169] Step S23: determining a second error of the logarithmic distribution error function based on the prior information, training expectation and training variance deviation corresponding to the computer room sample data;

[0170] Step S24: determining a combined error based on the first error and the second error, and optimizing network parameters of an initial temperature control parameter prediction neural network according to the combined error to obtain a temperature control parameter prediction neural network.

[0171] In step S21, the computer system processes the data in the calibration sample library. The calibration sample library includes computer room sample data of multiple computer rooms and prior information corresponding to the sample data of each computer room.

[0172] Assume that the initial temperature control parameter prediction neural network has a specific structure, such as an input layer, a hidden layer, and an output layer. For each computer room sample data in the calibration sample library, the computer system pre-processes it and loads it into the input layer of the neural network. Taking a computer room sample data as an example, the computer room facility parameters (such as the number of servers, power, etc.), temperature history control data (such as the historical set temperature of the air conditioner, cooling power, etc.) and temperature history control timing data (such as the equipment code sequence triggered by cooling, etc.) of the computer room are integrated and encoded to form a feature vector. Assuming that the dimension of this feature vector is n (for example, n=100), this n-dimensional vector is input into the input layer of the neural network, and the number of neurons in the input layer is n.

[0173] Once the data enters the input layer, forward propagation begins in the neural network. Data flows from the input layer to the hidden layer. Assume that there are m neurons in the hidden layer (for example, m=50), and the neurons in the hidden layer are connected by the weight matrix W and the bias vector b. For each neuron i (i=1,2,…,n) in the input layer and each neuron j (j=1,2,…,m) in the hidden layer, the input of neuron j in the hidden layer is is the weighted sum of the outputs of the input layer neurons plus the bias, i.e. ,in is the output of neuron i in the input layer, is an element in the weight matrix W. Then, the output of hidden layer neuron j is Calculated by activation function f, for example, using ReLU activation function .

[0174] After being processed by the hidden layer, the data continues to propagate to the output layer. The output layer outputs the corresponding results according to the prediction task. In this example, the output layer outputs the training confidence, training expectation, and training variance deviation related to the prediction of the temperature control parameters. Assume that the output layer has k=3 neurons, corresponding to the training confidence, training expectation, and training variance deviation respectively. The calculation method of the output layer neurons is similar to that of the hidden layer. It is also weighted summed by the weight matrix and the bias vector, and then the final output is obtained through the activation function (which may be a different activation function, such as the Sigmoid function used to map the training confidence to between 0 and 1). For example, for the confidence neuron, its output o1 may be 0.6 after calculation, indicating that the computer system has a 60% confidence level in the current predicted temperature control parameter; for the expectation neuron, the output o2 may be 21 (assuming that the temperature setting value is 21 degrees Celsius); for the variance deviation neuron, the output o3 may be 0.5, indicating that the variance deviation of the predicted temperature control parameter is 0.5.

[0175] In step S22, the information gain error function is used to measure the difference between the model prediction result and the prior information, while taking into account the training confidence of the prediction. The training confidence reflects the reliability assessment of the model on the prediction result.

[0176] Assume that for a certain computer room sample data, the temperature control parameter in its prior information is the temperature setting value Celsius, and the training confidence C=0.6 obtained from the neural network output, the expected temperature setting value Celsius (the desired temperature setting here is the result of the neural network output).

[0177] One possible form of the information gain error function is (This is just an example form, the actual information gain error function may be more complicated). Substituting the values ​​into .calculate ,but This first error It reflects the impact of the difference between the predicted temperature and the prior temperature on the information gain when considering the training confidence. If the training confidence is high and the difference between the predicted temperature and the prior temperature is small, the first error will be small; conversely, if the training confidence is low or the difference between the predicted temperature and the prior temperature is large, the first error will be large.

[0178] In step S23, the logarithmic distribution error function is used to measure the difference between the prior information, training expectation and training variance deviation. It is based on the assumption of logarithmic distribution and takes into account the relationship between the training expectation and training variance deviation of the prediction result and the prior information. Continuing with the above computer room sample data as an example, the temperature setting value in the prior information Celsius, the training desired temperature setpoint output by the neural network degrees Celsius, training variance bias σ = 0.5.

[0179] One possible form of the logarithmic error function is (This is just an example form). First calculate , Substituting into the formula, we get .calculate , , ,but This second error It reflects the error of the difference between the training expected temperature setting value and the training variance deviation and the prior temperature setting value under the logarithmic distribution assumption.

[0180] In step S24, the computer system calculates the first error and the second error A simple combination method is weighted summation. Assuming that the weight of the first error is w1=0.4 and the weight of the second error is w2=0.6, the combined error is .

[0181] Substituting the values ​​into the equation, we get .calculate , ,but This combined error combines the information gain error considering confidence and the error based on logarithmic distribution, and fully reflects the difference between the model prediction results and the prior information.

[0182] The computer system uses an optimization algorithm (such as stochastic gradient descent algorithm) to optimize the initial temperature control parameters based on the combined error to predict the network parameters (such as weights and biases) of the neural network. Assume that an element in the weight matrix W of the neural network , according to the stochastic gradient descent algorithm, the update formula is , where α is the learning rate (e.g. α=0.01).

[0183] calculate It is necessary to derive the error function according to the structure of the neural network and its specific form. This derivation process involves the chain rule, because the error E is obtained through multiple calculation steps in the forward propagation process. For example, for the weights from the hidden layer to the output layer The calculation requires backpropagation of the error from the output layer to the hidden layer, taking into account the activation function and connection relationship of each neuron.

[0184] By continuously updating the network parameters according to the combined error, the computer system gradually optimizes the initial temperature control parameter prediction neural network, so that after multiple iterations, the prediction results of the neural network are closer to the prior information, thereby obtaining the final temperature control parameter prediction neural network. This final neural network can more accurately predict the temperature control parameters of the computer room, providing a reliable basis for the effective control of the computer room temperature.

[0185] The present application embodiment provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present application.

[0186] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above model predictive control method of the computer room temperature is implemented.

Claims

1. A model predictive control method for computer room temperature, characterized in that: The method comprises: Acquire control evaluation factor data of the computer room to be controlled, the control evaluation factor data including observation results corresponding to the computer room to be controlled and multiple reference indicators respectively, the multiple reference indicators including multiple computer room facility parameters, multiple temperature history control data and multiple temperature history control time series data, the temperature history control data including temperature control parameters of the computer room to be controlled on multiple temperature control devices, the temperature history control time series data representing temperature control device control time series data obtained based on the control of the computer room to be controlled on the multiple temperature control devices for a target number of times before the current moment, wherein the temperature history control time series data includes device codes of the multiple temperature control devices; Performing feature representation on the observation results corresponding to each equipment room facility parameter and the temperature history control data to obtain a first feature array, performing feature representation on the time series observation results corresponding to each temperature history control time series data to obtain a second feature array, and generating a first fused feature array based on each first feature array and each second feature array; Loading the first fused feature array to a plurality of first task models respectively, performing feature extraction based on each first task model to obtain a first task model extraction array output by each first task model; Determine the completeness of the control evaluation factor data in each evaluation factor set in advance, and generate a primary data integrity array representation of the control evaluation factor data based on the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, wherein the primary data integrity array representation is used to identify blank content of the observation result corresponding to the reference indicator in the control evaluation factor data; Based on the primary data integrity array representation, the influence coefficient of each first task model is determined, based on the influence coefficient, the first task model extraction array is fused to obtain a primary blank influence fusion array, and based on the merging result of the primary blank influence fusion array and the first task model extraction array, the temperature control parameters of the computer room to be controlled on the set temperature control equipment are obtained.

2. The model predictive control method for computer room temperature according to claim 1, characterized in that: The temperature control parameters of the room to be controlled on the set temperature control device are obtained based on the merging result of the primary blank impact fusion array and the first task model extraction arrays, including: Combining the primary blank impact fusion array and each first task model extraction array to obtain a second fusion feature array; Loading the second fused feature array or a third fused feature array generated based on the second fused feature array into a deep dense network to obtain a mapping result of the deep dense network; The mapping results are loaded into the S-type nonlinear transformation component, the identity transformation component and the smooth ReLU transformation component respectively, and the confidence of the temperature control triggering of the control room to be controlled, the expectation of the temperature control parameter of the control room to be controlled and the variance deviation of the temperature control parameter of the control room to be controlled are inferred respectively; Based on the confidence level, the expectation and the variance deviation, a temperature control parameter on a temperature control device is determined and set.

3. The model predictive control method for the temperature of a computer room according to claim 2, characterized in that: Before loading the second fused feature array or the third fused feature array generated based on the second fused feature array into a deep dense network, the method further includes: Loading the second fused feature array to a plurality of second task models respectively, performing feature extraction based on each second task model to obtain a second task model extraction array output by each second task model; Based on the completeness of the control evaluation factor data in each evaluation factor set and the feature array set for each evaluation factor set, a secondary data integrity array representation of the control evaluation factor data is generated, wherein the secondary data integrity array representation is used to identify blank content of indicator combinations between reference indicators in the control evaluation factor data; Determine the influence coefficient of each second task model based on the secondary data integrity array representation, and fuse the extraction arrays of each second task model based on the influence coefficient to obtain a secondary blank influence fusion array; The secondary blank influence fusion array and the primary blank influence fusion array are combined to obtain a third fusion feature array.

4. The model predictive control method for computer room temperature according to claim 1, characterized in that: The step of generating a first fused feature array according to each of the first feature arrays and each of the second feature arrays includes: Combining the first feature arrays into a fourth fused feature array, and combining the second feature arrays into a fifth fused feature array; Determine the indicator combination relationship of the fourth fusion feature array based on the multiple indicator combination models, and obtain the first array representation of the output of each indicator combination model; Loading the fifth fused feature array into a sequence processing recursive model to obtain a second array representation of an output of the sequence processing recursive model; combining the first array representation and the second array representation into a first fused feature array; The determining of the integrity of the control evaluation factor data in each pre-set evaluation factor set, and generating a primary data integrity array representation of the control evaluation factor data according to the integrity of the control evaluation factor data in each evaluation factor set and a feature array set for each evaluation factor set, comprises: For an evaluation element set, determine the number of blank evaluation elements corresponding to the blank observation results in the evaluation element set, subtract the number of blank evaluation elements from the total number of evaluation elements in the evaluation element set to obtain the total number of complete evaluation elements, divide the total number of complete evaluation elements by the total number of evaluation elements to obtain the completeness of the evaluation element set; For each set of evaluation factors, the corresponding completeness is calculated; For each evaluation factor set, its corresponding integrity is multiplied by its set feature array to obtain the adjustment feature array corresponding to each evaluation factor set, and the adjustment feature arrays corresponding to each evaluation factor set are combined to obtain the primary data integrity array representation.

5. The model predictive control method for the temperature of a computer room according to claim 3, characterized in that: The step of generating a secondary data integrity array representation of the control evaluation factor data based on the integrity of the control evaluation factor data in each evaluation factor set and a feature array set for each evaluation factor set includes: For each pair of evaluation factor sets, according to the completeness of each evaluation factor set of the control evaluation factor data in the pair of evaluation factor sets, determine the completeness of the indicator combination of the control evaluation factor data in the pair of evaluation factor sets, and according to the feature array set for each evaluation factor set in the pair of evaluation factor sets, determine the feature array of the pair of evaluation factor sets; According to the indicator combination completeness and feature array of each pair of evaluation element sets, the secondary data integrity array representation of each pair of evaluation element sets is determined; the secondary data integrity array representation of each pair of evaluation element sets is combined to obtain the secondary data integrity array representation of the control evaluation element data.

6. The model predictive control method for the temperature of a computer room according to claim 5, characterized in that: The method of determining the completeness of the index combination of the control evaluation factor data in the pair of evaluation factor sets according to the completeness of each evaluation factor set of the control evaluation factor data in the pair of evaluation factor sets, and determining the feature array of the pair of evaluation factor sets according to the feature array set for each evaluation factor set in the pair of evaluation factor sets, comprises: determining the product of the completeness of each evaluation factor set of the control evaluation factor data in the pair of evaluation factor sets as the completeness of the index combination of the control evaluation factor data in the pair of evaluation factor sets; Determine the dot product of the feature arrays set for each evaluation element set in the pair of evaluation element sets as the feature array of the pair of evaluation element sets; Determining the secondary data integrity array representation of each pair of evaluation element sets according to the indicator combination integrity and feature array of each pair of evaluation element sets includes: The indicator combination completeness and feature array of each pair of evaluation factor sets are multiplied to obtain the secondary data integrity array representation of each pair of evaluation factor sets.

7. The model predictive control method for computer room temperature according to claim 2, characterized in that: The method is implemented by a temperature control parameter prediction neural network. Before obtaining the control evaluation factor data of the computer room to be controlled, the method also includes a training process of the temperature control parameter prediction neural network, including: Obtain a calibration sample library and a result verification sample library, wherein both the calibration sample library and the result verification sample library are training databases including sample data of computer rooms of multiple computer rooms and prior information corresponding to the sample data of each computer room, the computer rooms corresponding to the sample data of the computer rooms in the calibration sample library and the result verification sample library are computer rooms in different areas, the computer room sample data include observation results corresponding to the computer room and multiple computer room facility parameters, observation results corresponding to the computer room and multiple temperature history control data, and observation results corresponding to the computer room and multiple temperature history control time series data, the temperature history control data include temperature control parameters of the computer room on multiple temperature control devices, the temperature history control time series data represent temperature control device control time series data obtained based on the control of the computer room on the target number of times on the multiple temperature control devices before the current moment, and the prior information is the temperature control parameters of the computer room for the target temperature control device in a preset period; Debugging and obtaining a temperature control parameter prediction neural network based on the calibration sample library; The temperature control parameter prediction neural network is verified based on the result verification sample library, wherein obtaining the control evaluation factor data of the computer room to be controlled is performed when the temperature control parameter prediction neural network is based on the verification.

8. The model predictive control method for the temperature of a computer room according to claim 7, characterized in that: Before verifying the temperature control parameter prediction neural network based on the result verification sample library, the method further includes: Acquire a process verification sample library including computer room sample data of multiple computer rooms and prior information corresponding to each computer room sample data; The temperature control parameter prediction neural network is cyclically updated based on the calibration sample library, and after each update, the temperature control parameter prediction neural network in the calibration process is verified according to the process verification sample library to obtain the neural network debugging accuracy; After each update, it is determined whether the debugging accuracy of the neural network obtained after multiple consecutive updates has converged, wherein the temperature control parameter prediction neural network is verified based on the result verification sample library when it is determined that the debugging accuracy of the neural network obtained after multiple consecutive updates has converged.

9. The model predictive control method for the temperature of a computer room according to claim 7, characterized in that: The method of debugging and obtaining a temperature control parameter prediction neural network based on the calibration sample library includes: Loading the computer room sample data in the calibration sample library into the initial temperature control parameter prediction neural network to obtain the training confidence, training expectation and training variance deviation output by the initial temperature control parameter prediction neural network; Determine a first error of an information gain error function based on the training confidence and the prior information corresponding to the computer room sample data; Determine a second error of the logarithmic distribution error function based on the prior information corresponding to the computer room sample data, the training expectation and the training variance deviation; A combined error is determined based on the first error and the second error, and network parameters of the initial temperature control parameter prediction neural network are optimized according to the combined error to obtain a temperature control parameter prediction neural network.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images