Nonlinear model predictive control system for cooling unit system in data center

By adopting a nonlinear model predictive control system in the data center cooling unit system, and using dual control loops and prediction controllers, the problem of difficult disturbances in the frozen water flow control is solved, and the stability and control quality of the air supply temperature are improved. Save energy consumption.

CN117389146BActive Publication Date: 2024-11-15BEIJING UNIV OF CIVIL ENG & ARCHITECTURE +1
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202311482873.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-09
Publication Date
2024-11-15
Estimated Expiration
2043-11-09

AI Technical Summary

Technical Problem

In the data center cooling unit system, the refrigerated water flow control adopts a single loop structure, which cannot suppress frequent disturbances in time, resulting in fluctuations in the supply air temperature and affects the control quality.

Method used

The nonlinear model prediction control system is adopted, including the air supply volume control loop and the air supply temperature control loop, and the direct prediction control and cascade prediction control structure are used, combined with the predictive controller and the machine room temperature prediction model, and the dual control loop design and optimization goals are designed and optimized. Function, suppress disturbance and keep the air supply temperature stable.

Benefits of technology

Timely suppression of frequent disturbances is achieved, the control quality of the cooling unit system is improved, and the energy consumption of fan and refrigeration water flow is saved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117389146B_ABST
    Figure CN117389146B_ABST
Patent Text Reader

Abstract

A nonlinear model predictive control system for a cooling unit system in a data center, the nonlinear model predictive control system comprising an air supply volume control loop and an air supply temperature control loop; the air supply volume control loop adopts a direct predictive control structure, the control quantity is a fan frequency conversion control signal, and the controlled variable is the maximum cabinet inlet temperature; the air supply temperature control loop adopts a cascade predictive control structure; the output control quantity of the outer loop of the cascade predictive control structure is a chilled water flow set value, and the controlled variable of the outer loop of the cascade predictive control structure is the maximum cabinet inlet temperature; the output control quantity of the outer loop is a reference input of the inner loop of the cascade predictive control structure; the output control quantity of the inner loop of the cascade predictive control structure is a water valve control signal, and the controlled variable is the chilled water flow; the outer loops of the direct predictive control structure and the cascade predictive control structure both adopt predictive controllers; the predictive controller performs predictive control according to the data center cooling unit model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of data center cooling, in particular to a nonlinear model predictive control system for a data center cooling unit system. Background Art

[0002] Currently, data center energy-saving technologies mainly focus on reducing computing energy consumption, while cooling system energy consumption is not considered enough.

[0003] In addition, data center computer rooms need to be equipped with multiple cooling units, and the pipelines of each unit are usually connected in parallel. When the cabinet load changes, it is easy to cause pipeline pressure fluctuations, causing changes in the chilled water flow of other units, resulting in long-term fluctuations in the supply air temperature. The chilled water flow control of existing data center cooling units all adopts a single-loop structure, which cannot timely suppress frequent disturbances, making it difficult to maintain a stable supply air temperature, thus affecting the control quality of the cooling unit system. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a nonlinear model predictive control system for a data center cooling unit system, which can timely suppress frequently occurring disturbances so that the supply air temperature remains stable, thereby improving the control quality of the cooling unit system.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A nonlinear model predictive control system for a cooling unit system in a data center, wherein the nonlinear model predictive control system includes an air supply volume control loop and an air supply temperature control loop;

[0007] The air supply volume control loop adopts a direct predictive control structure, the control variable is the fan frequency conversion control signal, and the controlled variable is the maximum temperature of the cabinet inlet;

[0008] The air supply temperature control loop adopts a cascade predictive control structure;

[0009] The output control quantity of the outer loop of the cascade predictive control structure is the chilled water flow set value, and the controlled variable of the outer loop of the cascade predictive control structure is the maximum temperature of the cabinet inlet;

[0010] The output control quantity of the outer loop is the reference input of the inner loop of the cascade predictive control structure;

[0011] The output control quantity of the inner loop of the cascade predictive control structure is the water valve control signal, and the controlled variable is the chilled water flow rate;

[0012] The outer loops of the direct prediction control structure and the cascade prediction control structure both use a prediction controller;

[0013] The predictive controller performs predictive control according to a data center cooling unit model.

[0014] The beneficial effects of the present invention are:

[0015] Through the dual control loop design, combined with the predictive controller and the room temperature prediction model, frequent disturbances can be suppressed in time to keep the supply air temperature stable, thereby improving the control quality of the cooling unit system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The structural diagram of the cooling unit system for the data center;

[0017] Figure 2 is a system structure diagram of the present invention;

[0018] Figure 3 is the flow characteristic curve of the water valve component model;

[0019] Figure 4 It is the relationship diagram between supply air temperature and cold channel temperature;

[0020] Figure 5 It is a flowchart of the antecedent structure identification method;

[0021] Figure 6 is a flow chart of the CKF method;

[0022] Figure 7 This is the structural diagram of the neural network predictive controller;

[0023] Figure 8 Flowchart of adaptive adjustment algorithm for optimizing objective function weights;

[0024] Fig. 9 is the membership curve of the fuzzy linguistic variable (from left to right, corresponding to E, EC, and U);

[0025] Fig.10 is the geometric model of the computer room;

[0026] Fig.11 The original data (from left to right correspond to the cold aisle temperature and supply and return air temperature, cold aisle wind speed and supply air speed, cabinet inlet temperature, and each cabinet load);

[0027] Fig.12 This is the test result of the computer room temperature prediction model;

[0028] Fig.13 It is the cascade PID control structure diagram;

[0029] Fig.14It is a comparison chart (from left to right, it corresponds to the comparison of the maximum temperature of the cabinet inlet under different control methods, the comparison of the set value of the chilled water flow under different control methods, the comparison of the fan control signal under different control methods, and the adjustment curves of the three weight coefficients);

[0030] Fig.15 Predict the control system performance under different initial variable values ​​(from left to right corresponding to the first set of initial variable values ​​and the second set of initial variable values). DETAILED DESCRIPTION

[0031] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0032] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0033] Embodiment 1:

[0034] like Figures 1 to 15 As shown in Figure 1, the nonlinear model predictive control system of the data center cooling unit system is studied for the currently commonly used cold and hot aisle closed data center computer room system. The cooling unit system consists of a computer room chilled water air conditioner (Computer Room Air Handlers, CRAH), a raised floor, and a group of cold and hot aisles consisting of 2 rows of cabinets. The structural diagram is shown in Figure 1. Figure 1 As shown. The CRAH air conditioner is composed of fans, surface coolers, water valves and other components. The cabinets in the computer room adopt the common "face-to-face, back-to-back" form, that is, the air inlets of the two rows of cabinets face each other, and the cold air outlet is set in the middle channel to form a cold channel; the air outlets of the cabinet groups face each other to form a hot channel. The airflow organization form of the computer room is downward delivery and upward return, and the cold channel temperature control method is adopted, with the temperature at the air inlet of the cold channel as the controlled variable. In order to prevent the occurrence of cabinets with overheating risks, the controlled variable is further refined into the maximum cabinet inlet temperature (the highest inlet temperature of all cabinets).

[0035] Figure 1In the system, the CRAH air conditioner delivers cold air to the air supply static pressure box under the raised floor of the computer room. Due to the pressure difference, the cold air can flow into the cold channel through the floor grille. The cabinet then uses the fan to inhale the cold air and blows the hot air generated after heat dissipation to the hot channel; then the CRAH air conditioner recycles the hot air. In addition, the system has a cold and hot channel separation and cold channel isolation design to avoid mixing of cold air and air, thereby reducing the loss of cooling and air volume.

[0036] When the maximum cabinet inlet temperature setting value is a value that can ensure the safe operation of the server, the smaller the error between the cabinet inlet maximum temperature and its setting value, the more safe operation can be ensured. For the cooling unit system, the energy consumption generated by the water valve itself is relatively low, and the energy consumption of the cooling unit system is mainly generated by the variable frequency fan. Considering the overall energy saving of the cooling system, this embodiment incorporates the chilled water flow index into the optimization objective function, the purpose of which is to reduce the energy consumption of the chilled water pump of the refrigeration station system.

[0037] The control objectives of the nonlinear model predictive control system are to ensure safe operation of the data center, save fan energy consumption, and reduce the flow of chilled water.

[0038] The optimization goal of the nonlinear model predictive control system is to minimize the error between the maximum cabinet inlet temperature and the set value, while maximally saving fan energy consumption and reducing the chilled water flow.

[0039] The optimization objective function is:

[0040]

[0041] Where J(k) represents the prediction control optimization objective function; M is the prediction time domain; t1 is the starting time of the prediction time domain; T set represents the set value of the maximum temperature at the cabinet inlet; α represents the weight coefficient of the temperature tracking error term; f w [k] can reflect the change of water valve opening from the side, β represents f w [k] is the weight coefficient of the term; γ represents u fan The weight coefficient of the [k] term.

[0042] Based on the data center design specification: GB50174-2017, and combined with the actual conditions of the research object, the controlled variables and control quantity constraints are finally selected as follows:

[0043] Maximum temperature at the cabinet inlet: 18℃~27℃

[0044] Water valve control percentage: 0~100%

[0045] Fan control percentage: 0~100%

[0046] Air supply temperature: 17℃~30℃

[0047] Air supply speed: 1.5m / s~3m / s

[0048] Cabinet load: 1500~7000W / m 3

[0049] The structure of nonlinear model predictive control system is as follows Figure 2 As shown, Figure 2 x[k] in the equation represents the state variable of the system, i.e., the maximum cabinet inlet temperature T max [k]; (i=1,2,3,..,M) represents the predicted value of the system state variable output by the prediction model; x * [k+i] (i=1,2,3,..,M) represents the expected value of the state variable, that is, the set value T of the maximum temperature at the cabinet inlet set ; d[k] represents the disturbance of chilled water flow change caused by the pressure fluctuation of chilled water pipeline.

[0050] A nonlinear model predictive control system for a cooling unit system in a data center, wherein the nonlinear model predictive control system includes an air supply volume control loop and an air supply temperature control loop;

[0051] The air supply volume control loop adopts a direct predictive control structure, and the control quantity is the fan frequency conversion control signal u fan [k], the controlled variable is the maximum cabinet inlet temperature T max [k];

[0052] The air supply temperature control loop adopts a cascade predictive control structure;

[0053] The output control quantity of the outer loop of the cascade predictive control structure is the chilled water flow set value f in the optimization objective function. w [k], the controlled variable of the outer loop of the cascade predictive control structure is the maximum cabinet inlet temperature T max [k];

[0054] The output control quantity of the outer loop is the reference input of the inner loop of the cascade predictive control structure;

[0055] The output control quantity of the inner loop of the cascade predictive control structure is the water valve control signal u valve [k], the controlled variable is the chilled water flow rate;

[0056] The outer loops of the direct prediction control structure and the cascade prediction control structure both adopt a prediction controller.

[0057] The predictive controller performs predictive control according to a data center cooling unit model.

[0058] Through the dual control loop design, combined with the predictive controller and the room temperature prediction model, frequent disturbances can be suppressed in time to keep the supply air temperature stable, thereby improving the control quality of the cooling unit system.

[0059] The inner loop of the cascade predictive control structure adopts a PI controller.

[0060] Since the pressure fluctuation of the chilled water pipeline can easily cause long-term fluctuation of the supply air temperature, and cascade control is an effective method to deal with frequent disturbances, and predictive control can solve multi-variable and constrained control problems and improve the dynamic characteristics of the controlled object, the supply air temperature control loop absorbs the advantages of cascade control and predictive control and adopts a cascade predictive control structure.

[0061] The main and sub-controllers of traditional cascade control are both PID controllers, but they belong to the after-effect control method, which can easily lead to poor dynamic performance and cannot save energy while reducing the cabinet inlet temperature. In addition, for nonlinear systems, the setting and debugging of PID parameters are also very difficult. The predictive controller can improve the dynamic response quality and save energy based on multi-step prediction, rolling optimization and feedback correction. Therefore, the outer loop of the supply air temperature control loop uses a predictive controller instead of a PID controller to ensure the control accuracy of the system; because when there is a pulsating signal in the chilled water flow signal, the introduction of a differential link will amplify the disturbance and easily cause the water valve control to jitter, so the inner loop uses a PI controller.

[0062] Since the output u of the predictive controller fan [k] and f w [k] It is necessary to use various cooling system components to control the maximum cabinet inlet temperature T max [k] In view of the influence of the mechanism modeling methods of various system components and the maturity of the mechanism modeling methods, a data center cooling unit system component model is constructed by using the mechanism modeling method. The data center cooling unit system component model includes the surface cooler, water valve, and fan component models in the two control loops in the nonlinear model predictive control system.

[0063] The data center cooling unit model includes the data center cooling unit system component model and the computer room temperature prediction model;

[0064] The controller outputs the control quantity to the actuator, i.e. the surface cooler, water valve, and fan, which then acts to adjust the water flow and air flow so that the room temperature tracks the given value. The room temperature prediction model and the data center cooling unit system component model together constitute the controlled object model.

[0065] The nonlinear model predictive control system of the data center cooling unit system needs to construct a data center cooling unit system component model, and the data center cooling unit system component model is constructed by a mechanism modeling method, including a surface cooler component model, a water valve component model, and a fan component model;

[0066] The heat change per unit time of the surface cooler is the heat taken away by the air per unit time, minus the heat taken away by the chilled water per unit time. Therefore, it is necessary to construct the energy conservation equation of the surface cooler on the chilled water side and the air side.

[0067] The surface cooler component model is:

[0068] Difference equation on the chilled water side:

[0069]

[0070] Where, T wi (k) represents the chilled water temperature at the inlet of the condenser at time k (℃); T wo (k) represents the chilled water temperature at the outlet of the condenser at time k (℃); m w Indicates the chilled water mass flow rate (kg / s); M w Indicates the mass of chilled water stored inside the surface cooler (kg / s); T m represents the sampling period; c w Indicates the constant pressure specific heat capacity of chilled water (kg*℃);

[0071] The difference equation on the air side:

[0072]

[0073] Where, T ai (k) represents the inlet air temperature of the condenser at time k (℃); T ao (k) represents the outlet air temperature of the condenser at time k (℃); ξ represents the moisture separation coefficient; m a Indicates air mass flow rate (kg / s); M a Indicates the air mass stored inside the cooler (kg / s); c a Indicates the specific heat capacity of air at constant pressure (kg*℃);

[0074] Q(k) in the differential equation on the chilled water side and the differential equation on the air side represents the heat exchange between the air and the chilled water. The calculation formula of Q(k) is:

[0075] Q(k)=FhΔt m

[0076] Where F represents the heat exchange area of ​​the surface cooler (m 2 ), h represents the heat transfer coefficient, Δtm It represents the logarithmic mean temperature difference.

[0077] The water valve component model is:

[0078] G = 2.1 × 10 -10 L 7 -7.7×10 -8 L 6 +1.1×10 -5 L 5

[0079] -7.1×10 -4 L 4 +2.1×10 -2 L 3

[0080] -0.19L 2 +0.16L+0.14

[0081] Where G represents the relative flow rate of chilled water (kg / s); L represents the relative opening of the water valve (%).

[0082] In the construction of the water valve component model, a 7th-order polynomial model was constructed using the least squares curve fitting method based on the water valve data measured in the literature.

[0083] The fan component model is:

[0084] The calculation formula of the fan's air outlet speed is:

[0085] v f =v fr λ f

[0086] In the formula, v f Indicates the air outlet speed of the fan (m / s); v fr Indicates the rated wind speed of the fan (m / s); f Represents the fan control signal; the fan outlet speed calculation formula is established under the assumption that the fan control signal is 0~100%.

[0087] The fan control signal has a linear relationship with its air outlet speed and a cubic relationship with its energy consumption.

[0088] Fan energy consumption calculation formula:

[0089]

[0090] Where, P represents the energy consumption of the fan (kw); P r Indicates the energy consumption of the fan under rated conditions (kW); a i (i=0,1,2,3,…) represents polynomial factors.

[0091] Since the fan control signal and its energy consumption are in a cubic relationship, the polynomial factor a i Select 4. In addition, the constructed fan model was tested using the actual system operation data. According to the test results, the parameters of the fan model were calculated by the polynomial fitting method: v fr =1.5m / s, P r =6kW, a0=0.02, a1=2.432, a2=-1.9554, a3=0.583.

[0092] In addition to constructing a data center cooling unit system component model (i.e., a mathematical model of the main components), a computer room temperature prediction model is also required to achieve predictive control.

[0093] The data center cooling unit model also includes a computer room temperature prediction model.

[0094] In this embodiment, the fan speed and water valve opening are used as the control variables of the data center cooling system, both of which can significantly affect the maximum temperature of the cabinet inlet. In addition, the supply air temperature, cold channel temperature, cabinet load, and return air temperature also have an impact on the maximum temperature of the cabinet inlet.

[0095] Through experiments, it is found that the supply air temperature is highly correlated with the cold channel temperature. The results are as follows: Figure 4 As shown. Figure 4 It can be seen that the cold channel temperature can closely follow the supply air temperature. This is because the cold channel of the research object is close to the air supply outlet of the air conditioner, so the supply air temperature can approximately reflect the cold channel temperature. In addition, the server in the cabinet needs to inhale cold air from the cold channel for heat dissipation, and the hot air generated by heat dissipation is discharged into the hot channel. This hot air will eventually flow into the return air outlet of the air conditioner, so the return air temperature is closely related to the cabinet load.

[0096] The output of the outer loop predictive controller in the nonlinear model predictive control system structure is the reference input of the inner loop PI controller. Considering the rationality of the control structure design, the water valve opening control signal cannot be used as the input variable of the prediction model, and the chilled water flow set value is used instead.

[0097] Taking into account the correlation between the above variables, the input variables and output variables of the following prediction model are finally screened out.

[0098] Table 1 Input variables and output variables of the prediction model

[0099]

[0100] The input variables of the room temperature prediction model include the chilled water flow set value f w[k], fan frequency conversion control signal u fan [k], air supply temperature T in [k], maximum cabinet inlet temperature T max [k] and the load of each cabinet; the output variable of the room temperature prediction model is the maximum cabinet inlet temperature T at the next moment. max [k+1].

[0101] Due to the complexity of airflow organization and heat transfer process in data center computer room, mechanism modeling is difficult, so data-driven modeling is adopted, and training data is acquired by building a CFD simulation model of the computer room. Data center computer room usually contains multiple cold channels and hot channels, and multiple groups of cooling units need to be configured. In order to improve the efficiency of control system development, modeling needs to take into account both model accuracy and portability. Therefore, an improved TS fuzzy model of data center computer room temperature is proposed. Its antecedent structure is identified by the proposed IBAS-FCM algorithm, and the consequent parameters are identified by the CKF method.

[0102] The computer room temperature prediction model is a computer room temperature TS fuzzy model, which is a computer room temperature prediction model built based on data drive and the proposed improved TS fuzzy model.

[0103] TS fuzzy model description:

[0104] Assume that the sample data is S = {s(1), s(2), ..., s(k), ..., s(h)} (1 ≤ k ≤ h), where h represents the number of samples. The room temperature prediction model can be regarded as a TS fuzzy model of a nonlinear MISO system, i.e., y(k+1) = f(s(k))

[0105] Where s(k) represents the n-dimensional input variable at the current moment, s(k) = [s1(k), s2(k), …, s n (k)]; y(k+1) represents the output variable at the next moment; f(·) represents the nonlinear function of y(k+1) with respect to s(k).

[0106] The TS fuzzy model can be composed of l fuzzy rules, and the i-th fuzzy rule can be expressed as:

[0107] R i :If s1(k)is A i1 and s2(k)is A i2 and...ands n (k)is A in ,

[0108]

[0109] In the formula, R irepresents the i-th fuzzy implied conditional sentence, where the If statement represents the rule antecedent and the Then statement represents the rule consequent; s j (k) represents the j-th dimension input variable of the i-th rule; A ij represents the j-dimensional fuzzy subspace of the i-th rule; y i (k+1) The linear output of the ith rule; a i represents the set of consequent parameters of the i-th rule, a i =[a i0 ,a i1 ,...,a in ].

[0110] By using the weighted average method, we can get:

[0111] The TS fuzzy model output formula of the room temperature TS fuzzy model is:

[0112]

[0113] In the formula, β ik represents the applicability of s(k) to the i-th rule, y i (k+1) represents the linear output of the i-th rule; a i =[a i0 ,a i1 ,...,a in ], a i Represents the set of consequent parameters of the i-th rule;

[0114] β ik It can be obtained by the product method: Where μ(·) represents the jth membership function of the ith rule, which can determine the fuzzy subspace A ij θ ij represents the parameter of μ(·).

[0115] Through experiments, a Gaussian membership function is selected, and its expression is: In the formula, θ ij =[c ij ,σ ij ];c ij represents the center of the membership function; σ ij Indicates the width of the membership function.

[0116] In the TS fuzzy model, the antecedent parameter set θ to be identified is [θ ij ] l×n , the set of subsequent parameters to be identified a=[a1,a2,...,a l ].

[0117] The antecedent structure identification method of the room temperature TS fuzzy model is as follows: by taking the objective function of FCM as the fitness function of the BAS algorithm, taking the cluster center as the beetle position in the BAS algorithm, and continuously iterating to find the optimal beetle position x t , so that the clustering objective function is minimized.

[0118] Traditional TS fuzzy model usually uses FCM algorithm for antecedent structure identification. However, this algorithm is sensitive to the initial cluster center and is easy to fall into local optimum. Most of the existing results use particle swarm optimization algorithm (PSO) to optimize FCM, namely PSO-FCM algorithm, but PSO algorithm has the problem of large amount of calculation and long convergence time. Compared with PSO algorithm, beetle antennae search (BAS) algorithm requires less population and has the advantage of small amount of calculation, but traditional BAS algorithm itself is easy to fall into local optimum. Therefore, this paper improves BAS algorithm and proposes an improved beetle antennae search algorithm (IBAS). After calculating the positions of the left and right antennae of beetle, arithmetic crossover operator and Gaussian mutation operator are introduced to make the algorithm jump out of local optimum. IBAS algorithm is used to optimize FCM, namely IBAS-FCM algorithm, so as to realize antecedent structure identification.

[0119] The improvements of the IBAS algorithm compared to the BAS algorithm are as follows:

[0120] 1) In the IBAS algorithm, the arithmetic crossover operator is introduced to linearly combine two individuals to generate two new individuals.

[0121] 2) Introduce the Gaussian mutation operator and use Gaussian distribution random numbers to change the position of an individual.

[0122] 3) After the arithmetic crossover operator and Gaussian mutation operator are used, the position of the newly generated antenna is Evaluate if these new antenna positions are better than and Then a replacement is made, so that the algorithm can jump out of the local optimum.

[0123] The constraints of the membership matrix U are: ||U t+1 -U t ||<ε

[0124] Where U t+1 represents the membership matrix of the t+1th iteration; U t represents the membership matrix of the tth iteration; ε represents the convergence accuracy.

[0125] The IBAS-FCM algorithm uses the cluster center C as the position of the longicorn, that is:

[0126] The preceding structure identification method comprises the following steps:

[0127] 1) Initialization parameters;

[0128] Includes parameters of the FCM and IBAS algorithms, including: ① number of clusters l; ② search step d s ; ③ Attenuation factor eta; ④ Step factor r; ⑤ Crossover probability P c ⑥ Mutation probability P m ; ⑦ Crossover factor α; ⑧ Maximum number of iterations t max ;

[0129] 2) Randomly generate the initial position of the longicorn;

[0130] 3) Calculate the search direction according to the following two formulas and the induction length p t ;

[0131]

[0132]

[0133] In the above two formulas, rands(·) represents the generated n-dimensional random number; p t represents the sensing length of the antennae of the longicorn beetle at the tth iteration; d s t represents the search step size at the tth iteration; d s t-1 represents the search step size at the t-1th iteration.

[0134] 4) By searching direction and the induction length p t Update antenna position and x r t ;

[0135] The antenna position and The update formula is: In the formula, x t is the antenna position of the longicorn beetle at the tth iteration, is the search direction, p t is the sensing length, is the position of the left antenna at the tth iteration, is the position of the right antenna at the tth iteration;

[0136] In step 4), the antenna position and Perform optimization operations;

[0137] The optimization operation is based on the antenna position and A crossover operation or a mutation operation is performed to obtain a new antenna position. If the new antenna position is better than the antenna position, the antenna position is replaced by the new antenna position.

[0138] The crossover operation is to satisfy the crossover condition by using the following formula according to the antenna position and Execute the crossover operator:

[0139]

[0140] In the formula, α represents the crossover factor, is the left and right antenna positions after arithmetic crossing, that is, the new antenna positions.

[0141] The crossover operation introduces the arithmetic crossover operator to linearly combine two individuals to produce two new individuals. Specifically, the position of the left antenna at the tth iteration and the position of the right antenna at iteration t Perform the arithmetic crossover operator operation to obtain the left and right antenna positions after arithmetic crossover and

[0142] The crossover condition is that the random number rand ≤ the crossover probability P c The random number rand range is [0,1].

[0143] The mutation operation is to satisfy the mutation condition by using the following formula according to the antenna position and Execute the mutation operator:

[0144]

[0145] In the formula, x best represents the global optimal position of the longicorn; Gaussian(1,1) represents a Gaussian distributed random number with a mean and variance of 1. and They are the positions of the left and right antennae after Gaussian mutation.

[0146] The mutation operation is a Gaussian mutation operator operation. The Gaussian mutation operator is introduced to change the position of an individual using Gaussian distribution random numbers.

[0147] The mutation condition is that the random number rand ≤ mutation probability P m

[0148] After the arithmetic crossover operator and Gaussian mutation operator are used, the positions of the newly generated antennae are Evaluate if these new antenna positions are better than and Then a replacement is made, so that the algorithm can jump out of the local optimum.

[0149] 5) Get the antenna position and The corresponding fitness;

[0150] Tentacle location and The corresponding fitness acquisition method is: calculate the antenna position according to the membership update formula and The corresponding membership matrix is ​​used to calculate the antenna position according to the objective function of FCM clustering. and The corresponding fitness;

[0151] The membership update formula is: In the formula, Indicates the sample point s(k) of the t+1th iteration for c i The membership degree of jk Represents the sample point s(k) to the cluster center c j The Euclidean distance.

[0152] The objective function of the FCM clustering is: In the formula, C represents the cluster center, C=[c1,c2,...,c l ]; m represents the fuzzy weighted index, which is used to control the fuzziness of the membership matrix U, U = [u ik ] l×h , where u ik Indicates the sample point s(k) for c i The membership degree of u ik satisfy d ik Represents the sample point s(k) to the cluster center c i The Euclidean distance, d ik =||s(k)-c i ||2.

[0153] 6) Update the position x of the beetle antenna according to the following formula: t , and update the global optimal position of the longicorn according to the fitness;

[0154]

[0155] In the formula, x t-1 is the antenna position of the longicorn beetle in the t-1th iteration; fs (·) represents fitness function; sign(·) represents sign function;

[0156] sign(·) can be expressed as:

[0157] 6.1) Obtain the fitness corresponding to the position of the antennae of the longicorn beetle;

[0158] 6.2) The antenna position of the longhorn beetle and the global optimal position of the longhorn beetle with lower fitness are updated as the global optimal position of the longhorn beetle;

[0159] For example, the position of the antennae of the longhorn beetle is x t The fitness of the beetle is less than the global optimal position x best The global optimal position of the longhorn beetle is the position of the beetle antenna x t , that is, x best =x t .

[0160] The method for obtaining the fitness corresponding to the antenna position of the longicorn beetle is the same as step 5).

[0161] 7) Determine whether to perform the next iteration. If the next iteration is performed, t is incremented by 1 and jumps to step 3);

[0162] Determine whether t satisfies t≤t max If yes, proceed to the next iteration, t=t+1 and jump to step 3), otherwise terminate the algorithm;

[0163] 8) Obtain the globally optimal rule antecedent parameters through calculation;

[0164] After outputting the global optimal cluster center, the width σ of the membership function is calculated by the following formula ij , and obtain the globally optimal rule antecedent parameters:

[0165] Common methods for the identification of the posterior parameters of nonlinear system models include EKF and Unscented Kalman Filter (UKF). However, the use of EKF for nonlinear system state estimation will introduce linearization errors, resulting in low estimation accuracy and poor stability. UKF can also estimate the state of nonlinear systems, but as the system dimension increases, the scaling ratio may be negative, which may easily cause the error covariance matrix to be non-positive definite, resulting in a decrease in filtering accuracy. The cubature Kalman filter (CKF) method is based on the third-order spherical radial criterion and uses a cubature point set to approximate the posterior distribution of the system state. It can be regarded as an effective posterior parameter identification method. Compared with EKF, CKF reduces the linearization error and makes the system state estimation have third-order accuracy. Compared with UKF, CKF can ensure that the error covariance matrix is ​​positive definite regardless of the system dimension, and has better estimation accuracy and stability.

[0166] The subsequent parameter identification method of the TS fuzzy model of the room temperature is the CKF method, that is, the cubature Kalman filter method.

[0167] The nonlinear system used for parameter estimation in the CKF method is:

[0168] a(k)=f h (a(k-1))+w(k-1)

[0169] z(k)=h(a(k))+v(k)

[0170] Where a(k) represents the system state at time k (0≤k≤h), which is the subsequent parameter of the TS fuzzy model of the room temperature to be identified; w(k) and v(k) represent independent Gaussian white noises with zero mean and covariances of Q(k) and R(k) respectively; z(k) represents the system measurement value at time k; f h (·) and h(·) represent the system equation and the measurement equation respectively. h(·) is obtained by the TS fuzzy model output formula of the room temperature TS fuzzy model. h (·) is obtained by the following formula:

[0171] f h (a(k-1))=ψ(k-1)a(k-1)

[0172] Where Ψ is a constant matrix, and a(k-1) represents the system state at time k-1 (0≤k-1≤h).

[0173] Ψ relates the parameters to each other, and if all parameters are freely adjustable, then Ψ is an identity matrix.

[0174] The expression of a(k) is:

[0175] like Figure 6 As shown, the CKF method includes the following steps:

[0176] 1) Initialize system state estimation And the error covariance P0:

[0177] 2) Time update

[0178] ① Assume that the posterior density function p(a(k)) follows a Gaussian distribution at time k-1, that is, Then, the error covariance P(k-1|k-1) at time k-1 is decomposed by Cholesky, and the decomposed matrix is: S h (k-1|k-1)=chol(P(k-1|k-1))

[0179] Where chol(·) represents the function for calculating Cholesky decomposition.

[0180] ② Apply the third-order spherical radial criterion to calculate the volume point set: In the formula, r h is the dimension of the consequent parameter estimation system, r h =l×n;m h Represents the number of volume points, which is the dimension r h twice;[1] i represents the i-th point of [1], [1]∈R r For r h r in the dimensional state space h The set of group symmetry points, namely:

[0181]

[0182] ③Calculate m at time k-1 h Volume points:

[0183] ④ Calculate the propagation volume points from k-1 to k through the state equation: A i (k|k-1)=f h (A i (k-1|k-1))

[0184] ⑤ According to the volume points, calculate the state prediction value and its error covariance from k-1 to k:

[0185]

[0186]

[0187] 3) Measurement update

[0188] ① Perform Cholesky decomposition on the error covariance P(k|k-1): S h (k|k-1)=chol(P(k|k-1))

[0189] ② Calculate m from time k-1 to time k h Volume points, namely:

[0190] ③ Calculate the propagation volume point through the measurement equation, that is: Y i (k|k-1)=h(A i ′(k|k-1))

[0191] ④ Calculate the error covariance of the measured predicted values ​​and their autocorrelation and cross-correlation:

[0192]

[0193]

[0194]

[0195] 4) Filter update

[0196] ① Calculate the Kalman filter gain at time k:

[0197] ② Calculate the state estimate and its error covariance at time k:

[0198]

[0199] P(k|k)=P(k|k-1)-K(k)P yy (k|k-1)K T (k)

[0200] The prediction controller is a neural network prediction controller, and the prediction control method of the neural network prediction controller is a neural network nonlinear rolling optimization method.

[0201] For a time-varying nonlinear MIMO system, let its prediction model (i.e., system equation) be:

[0202] x[k+1]=f(x[k],u[k],k)

[0203] x[0]=x0~P(x0)

[0204] Where x[k] represents the n-dimensional state vector measured at time k, x[k] = [x1(k), x2(k), …, x n (k)]; u[k] represents the m-dimensional control vector at time k, u[k] = [u1(k),u2(k),…,um (k)]; x[k+1] represents the state vector at time k+1, which is determined by x[k] and u[k] through the nonlinear function f(·); x0 is the initial state quantity, which obeys the probability distribution of P(x0). Here, the prediction model is the room temperature prediction model.

[0205] Previous studies have shown that a single hidden layer neural network can fit any nonlinear relationship through adaptive learning. Therefore, a three-layer feedforward neural network is selected as the feedback optimization controller to achieve online optimization of the control sequence of the nonlinear system. The basic expression of the neural network controller (i.e., the controller equation) is: u[k] = g(x[k], x * [k+1],W)

[0206] In the formula, x * [k+1] represents the expected value of the state at the next moment; W represents the weight matrix of the neural network controller (including the threshold); the input variables of the neural network controller are x[k], x * [k+1] and -1 corresponding to the threshold, the output variable is u[k].

[0207] The neural network predictive control rolling optimization algorithm can solve the optimal control quantity sequence u[k], which minimizes the system optimization objective function J under the constraints given by the system equation and the controller equation. J can be expressed as:

[0208]

[0209] Where M is the prediction time domain; represents the performance index of the terminal in the prediction time domain; L(x[k],u[k],k) represents the performance index of the process. Here, the optimization objective function J is expressed as the optimization objective function, that is, formula (1), and L(x[k],u[k],k) is the right side of the optimization objective function (that is, formula (1)).

[0210] In order to solve the functional extreme value of J, according to the application theorem of variational method in optimal control, the Lagrange multiplier method is adopted, the unknown n-dimensional Lagrange multiplier vector λ[k] and m-dimensional Lagrange multiplier vector q[k] are introduced, and the augmented functional J of J is constructed. a :

[0211]

[0212] Where f(k) and g(k,W) represent the right-hand side of the system equation and the controller equation, respectively; x[t0] represents the final state value of the previous rolling optimization, that is, the system sampling value, which is fed back to the controller and used as the initial state value x[t1] of this optimization.

[0213] Through the augmented function Ja The objective function extreme value with constraints (system equations and controller equations) can be transformed into the functional extreme value without constraints.

[0214] Introducing Hamiltonian function

[0215] J a Can be rewritten as:

[0216] Since the target here is the process control system, the optimization objective function does not include the performance index at the terminal moment, that is,

[0217] Let x[k], u[k] and W be relative to the optimal value x * [k]、u * [k] and W * The variations of are δx[k], δu[k] and δW respectively, then the functional J caused by these variations is a Variation of δJ a It can be expressed as:

[0218]

[0219] For any δx[k], δu[k], and δW, the functional J a The necessary condition to reach the extreme value is its variation δJ a = 0. Since x[t0] is the initial value of the state, it is usually a constant, so δx[t0] = 0, J a To reach the extreme value of the functional, the following conditions must be met (Hamiltonian canonical equation, transversality condition, extreme value condition).

[0220] Then the Hamiltonian canonical (adjoint) equation can be derived as:

[0221]

[0222]

[0223] The transversality condition is:

[0224] The extreme value conditions are:

[0225] Neural network predictive controller structure:

[0226] The input variables of the neural network prediction controller include air supply temperature, cabinet load (load of 2 rows of cabinets), cabinet inlet maximum temperature, cabinet inlet maximum temperature setting value, and -1 related to the neural network threshold, wherein -1 is introduced to complete the weight (threshold) value matrix conversion;

[0227] The output variables of the neural network prediction controller are control quantities, including a chilled water flow set value and a fan frequency conversion control signal.

[0228] Here, it is taken as an example that each row of cabinet groups contains 4 cabinets, that is, 2 rows of cabinet groups contain 8 cabinets. The input variables and output variables of the controller are shown in Table 2.

[0229] Table 2 Controller input variables and output variables

[0230]

[0231] Here, an empirical formula is used to determine the number of neurons in the hidden layer of the neural network controller: In the formula, l h represents the number of hidden layer nodes; l n Indicates the number of input layer nodes; l o represents the number of output layer nodes; a represents a constant, and its value range is [1,10]. The number of hidden layer neurons calculated by the empirical formula is in the range of [5,14]. Then, through simulation performance comparison, it is determined that the number of hidden layer neurons of the neural network predictive controller is 6.

[0232] The weight coefficient of the optimization objective function is obtained by optimizing the objective weight adaptive module;

[0233] The weight coefficient is obtained by optimizing the target weight adaptive module, and the weight coefficient is applied to the optimization target function of the neural network predictive controller;

[0234] The workflow of the optimization target weight adaptive module is as follows:

[0235] s01: Calculate the first deviation value at time k according to the maximum temperature at the cabinet inlet at time k and the set value of the maximum temperature at the cabinet inlet, and calculate the second deviation value at time k according to the maximum temperature at the cabinet inlet at time k and the maximum temperature at the cabinet inlet at time k-1;

[0236] e[k]=T set -T max [k]

[0237] ΔT[k]=T max [k-1]-T max [k]

[0238] T max [k] is the maximum cabinet inlet temperature at time k, T set is the maximum temperature setting value of the cabinet inlet, T max [k-1] is the maximum cabinet inlet temperature at time k-1, e[k] is the first deviation value at time k, and ΔT[k] is the second deviation value at time k;

[0239] s02: Calculate the weight coefficient α[k] of the tracking error term at time k according to the first deviation value at time k;

[0240] α[k]=α[k-1]+K pf (e[k]-e[k-1])

[0241] α[k-1] is the weight coefficient of the tracking error term at time k-1, e[k-1] is the first deviation value at time k-1, K pf represents the proportionality coefficient, and K is set by trial and error. pf Set to 0.48;

[0242] s03: Determine whether the first deviation value at time k is within the domain of the first deviation value. If so, use the fuzzy system to calculate the weight coefficient β[k] of the chilled water flow set value item at time k according to the first deviation value at time k and the second deviation value at time k. Otherwise, the weight coefficient β[k] of the chilled water flow set value item at time k is the weight coefficient β[k-1] of the chilled water flow set value item at time k-1.

[0243] s04: The weight coefficient γ[k] of the fan variable frequency control signal item at time k is calculated based on the weight coefficient α[k] of the tracking error item at time k and the weight coefficient β[k] of the chilled water flow setting value item at time k.

[0244] γ[k]=1-α[k]-β[k]

[0245] The fuzzy system is used to calculate the weight coefficient β[k] of the chilled water flow set value item at time k:

[0246] The first deviation value e[k] at time k and the second deviation value ΔT[k] at time k are input to the fuzzy system, and the fuzzy system outputs Δβ[k].

[0247] β[k]=β[k-1]+Δβ[k]

[0248] The optimization target weight adaptive module adopts the optimization target function weight adaptive adjustment algorithm. Through certain rules, it obtains the weight coefficient α[k] of the tracking error term at time k, the weight coefficient β[k] of the chilled water flow set value term, and the weight coefficient γ[k] of the fan variable frequency control signal term.

[0249] Fuzzy domain:

[0250] The fuzzy system is designed by incremental output. The input of the fuzzy system is the first deviation value e and the second deviation value ΔT, and the corresponding fuzzy linguistic variables are E and EC; the output is the correction value Δβ of the weight coefficient β of the chilled water flow set value item, and the corresponding fuzzy linguistic variable is U.

[0251] In the fuzzification process, it is necessary to determine the domain of e, ΔT, Δβ, E, EC, and U. Through a large number of simulations, the domains of the above variables are determined as shown in Table 3.

[0252] Table 3 Domains of various variables

[0253]

[0254] After determining the basic domain and fuzzy domain, the basic domain of each variable is scaled to convert it into the fuzzy domain. e , K ec , K u , it is possible to transform the basic domain of e, ΔT, Δβ into the fuzzy domain of E, EC, U. Finally, we can obtain K e =4 / 0.04=100, K ec =2 / 0.002=1000, K u =0.03 / 4=0.0075.

[0255] The domains of the fuzzy linguistic variables E, EC, and U are [-4,4], [-2,2], and [-4,4], respectively, and the corresponding scaling factors K e , K ec , K u They are 100, 1000 and 0.0075 respectively.

[0256] The domains of the correction amounts of the first deviation value, the second deviation value and the weight coefficient of the chilled water flow rate setting value item are [-0.04, 0.04], [-0.002, 0.002] and [-0.03, 0.03] respectively.

[0257] Fuzzy linguistic value and membership function:

[0258] The specification GB / T 19413-2010 "Unitary air conditioners for computer and data processing rooms" requires that the temperature control accuracy of the air conditioner in the data center room is ±1°C. Considering the fact that the domain becomes more refined after the scale transformation, the number of fuzzy language values ​​of the fuzzy language variables E, EC, and U are 4, 5, and 5 respectively for the consideration of real-time performance and control effect. Therefore, the fuzzy subset of E can be expressed as [negative large (NB), negative small (NS), positive small (PS), positive large (PB)]; the fuzzy subset of EC can be expressed as [negative large (NB), negative small (NS), zero (Z), positive small (PS), positive large (PB)]; the fuzzy subset of U can be expressed as [negative large (NB), negative small (NS), zero (Z), positive small (PS), positive large (PB)].

[0259] In view of the fact that the triangular (trapezoidal) function is easy to apply and calculate, the fuzzy linguistic variables E and EC adopt the trapezoidal membership function, and the fuzzy linguistic variable U adopts the triangular membership function.

[0260] Fuzzy rules:

[0261] Through a large number of simulations, the influence of different weight coefficient combinations on the system regulation effect is summarized, and the following experience is obtained: the larger α is, the smaller the steady-state error of the system is; the larger β is, the greater the fan energy consumption is; the larger γ is, the smaller the fan energy consumption is; a larger β can eliminate T max Fluctuations around its set value, and a larger γ cannot eliminate T max Fluctuations around its set value.

[0262] In summary, the weight coefficient adjustment strategy of the optimization target weight adaptive module is: when the first deviation value e is large, the overall performance of the system focuses more on reducing the deviation. At this time, α increases, β decreases, and γ remains unchanged, so as to improve the response speed of the system; when the first deviation value e is small, the overall performance of the system focuses more on saving fan energy consumption. At this time, α remains unchanged, β decreases, and γ increases; when the first deviation value e is small but the maximum temperature of the cabinet inlet fluctuates around its set value, α remains unchanged, β increases, and γ decreases.

[0263] Combining the weight coefficient adjustment experience and adjustment strategy, the fuzzy control rule table is finally established as shown in Table 4.

[0264] Table 4 Fuzzy control rules table

[0265]

[0266]

[0267] In order to enhance the adaptive ability of the control strategy and improve the control effect, an optimization target weight adaptive module is constructed based on fuzzy logic to adjust the optimization target weights α, β, and γ online.

[0268] Design a weight adaptive module based on fuzzy logic to dynamically and adaptively adjust the weight coefficients in the optimization objective function;

[0269] The neural network nonlinear rolling optimization method comprises the following steps:

[0270] (1) Initialize the weights of the neural network predictive controller, set the weights of the neural network predictive controller to be random numbers in the range of [-1,1], select the prediction time domain M and learning rate μ of the neural network predictive controller

[0271] The prediction time domain M is selected based on the principle that the prediction time domain should cover the entire transient range of the controlled object, and the learning rate μ is selected based on the trial and error method;

[0272] (2) Assume that the current time k = t1, set the expected value x*[k] (k = t1, ..., t1 + M-1) at each time in the prediction time domain, and use the actual system output at the previous time as the initial state value x[t0] of the rolling optimization at the current time;

[0273] k=t1, that is, the current time k is set as the starting time t1 of the prediction time domain;

[0274] (3) Setting weight coefficients of the optimization objective function, for example, calculating a first deviation value e[t1] by Tmax[t1] and Tset, calculating a second deviation value ΔT[t1] by Tmax[k-1] and Tmax[k], obtaining weight coefficients α[t1], β[t1], and γ[t1] by optimizing the objective function weight adaptive module, and applying these three weight coefficients to the optimization objective function of the neural network nonlinear rolling optimization method;

[0275] (4) The state quantity x[t1] = {Tin[t1], R1[t1], R2[t1], R3[t1], R4[t1], R5[t1], R6[t1], R7[t1], R8[t1], Tmax[t1]} of the system at time t1, the set value Tset[t1+1] of the maximum temperature at the cabinet inlet at time t1+1, and the threshold -1 are applied to the neural network predictive controller to obtain the control quantity u[t1] = {ufan[t1], fw[t1]}; then x[t1] and u[t1] are applied to the room temperature prediction model to obtain the predicted value of the state quantity at the next time.

[0276] (5) As in step (4), the predicted value of the state quantity at time t1+i is The set value Tset[t1+i+1] of the maximum temperature at the cabinet inlet at time t1+i+1 and the threshold -1 act on the neural network predictive controller to obtain the control quantity u[t1+i], where i represents the current prediction step number (i=1,2,3,…,M-1); then u[t1+i] acts on the room temperature prediction model to obtain the predicted value of the state quantity at time t1+i+1 And store all calculation results;

[0277] (6) Based on the results of steps (4) and (5), calculate the Lagrange multipliers λ[k], q[k] (k = t1+M-1, ..., t1+1, t1) from back to front according to the Hamiltonian canonical equation;

[0278] That is, from k = t1 + M-1 to k = t1, calculate λ[k], q[k] (k = t1 + M-1, ..., t1 + 1, t1);

[0279] (7) adjusting the weights of the neural network predictive controller based on the stochastic gradient descent method according to q[k] calculated in step (6);

[0280] Based on the stochastic gradient descent method, the following two equations can be inferred and used to adjust the weights of the neural network controller;

[0281]

[0282] W=W+ΔW

[0283] Where μ is the learning rate; W represents the weight matrix (including the threshold) of the neural network predictive controller.

[0284] (8) Repeat steps (4) to (7) and continuously adjust the weights of the neural network until the weights no longer change, that is, ΔW = 0; apply the control quantity at this time to the cooling system of the data center to obtain the maximum cabinet inlet temperature T at the next moment. max [t1+1];

[0285] Specifically, u fan [t1] acts on the fan of the variable air volume control loop to control the air volume; f w [t1] acts on the inner loop PI controller of the air supply temperature control loop, according to the actual value of the chilled water flow and f w [t1], calculate the linear combination of the proportional term and the integral term to obtain the control signal u of the water valve valve [t1], in order to suppress the supply air temperature fluctuation caused by the frequent changes in the chilled water flow; by controlling the supply air volume and supply air temperature, the maximum cabinet inlet temperature T at the next moment is finally obtained. max [t1+1].

[0286] (9) Update the time, set k = t1 + 1, repeat steps (2) to (8), and perform a new round of optimization until the entire control process is completed.

[0287] For the neural network nonlinear rolling optimization method, assuming that the dimension of the state quantity is n, the dimension of the control quantity is m, and the number of weights of the neural network predictive controller is n w , the prediction time domain is M, and the optimization needs to be performed (2n(n+m)+n w (m+2))M steps of calculation, taking about (n+m)M+2n w storage units, the storage space required by the algorithm is linearly related to the state dimension, control dimension, and controller weight number; among the currently commonly used nonlinear optimization algorithms, the computational complexity of the dynamic programming algorithm is exponentially related to the system dimension, and the time complexity of the sequential quadratic programming algorithm (SQP) is O(n3 ), the space complexity is O(n 2 ). Therefore, when solving high-dimensional nonlinear optimization problems, the neural network nonlinear rolling optimization method has small computational complexity, moderate storage space occupation, and is easy to implement using a low-cost field controller.

[0288] Commonly used nonlinear optimization algorithms, including dynamic programming and quadratic programming, have the problems of large amount of calculation, large storage space occupation, and difficulty in engineering implementation. Therefore, a neural network nonlinear rolling optimization method is proposed for Figure 2 The neural network predictive controller in the paper is used, and the neural network structure design and corresponding rolling optimization mechanism design are carried out according to the control requirements of the cooling unit in the data center. The nonlinear system optimization solution is realized under the premise of small calculation amount and moderate storage space occupation.

[0289] The parameters of the inner loop PI controller are set using a commonly used particle swarm optimization (PSO) method to suppress the frequent disturbances of the chilled water flow rate changes and ensure the rapidity of the system response.

[0290] Experimental research:

[0291] A simulation experiment was conducted on the cooling unit system of a data center in Beijing to verify the effectiveness of the proposed method. Since the project is currently under construction, the system operation data cannot be obtained; in addition, the actual operation data is usually maintained at a fixed condition, making it difficult to capture the overall dynamic characteristics of the system. To this end, a numerical simulation method was used to establish a CFD model of the data center computer room; on this basis, a simulation experiment was conducted on the constructed model using the proposed predictive control strategy.

[0292] Prediction model training and testing

[0293] Data collection and preprocessing

[0294] The transient CFD numerical simulation method is used to establish the CFD model of the computer room, simulate the temperature flow field of the computer room, and provide operating data and simulation environment for computer room temperature prediction and unit energy-saving optimization.

[0295] The construction of the computer room CFD model is divided into three parts, including pre-processing, numerical solution and post-processing. The pre-processing is implemented using Icepak software, which has a rich geometric component library and can quickly build a geometric model; the numerical solution is implemented using Fluent software, which allows users to customize control equations, material properties, etc., and can be used to perform numerical simulation of the data center computer room; the post-processing is implemented using CFD Post software, which can observe and analyze the calculation results of transient simulation.

[0296] The geometric model of the data center room is as follows: Fig.10The basic dimensions of the computer room are as follows:

[0297] 1) The size of the machine room is 6.9m×5.6m×5m;

[0298] 2) The size of the air conditioner is 1.5m×2m×2m;

[0299] 3) The dimensions of the air outlet and return air outlet of the air conditioner are 0.6×0.7m;

[0300] 4) The size of the cabinet is 0.6m×1m×2m;

[0301] 5) The size of the grid floor is 0.6m×0.6m;

[0302] 6) The width of the cold channel is 1.2m;

[0303] 7) The distance between the cabinet group and the side wall is 1.2m;

[0304] 8) The height of the raised floor is 0.7m.

[0305] Through the open-loop step response study, the time constant of the data center cooling system is calculated to be 34s. Based on the principle that the sampling period is generally less than or equal to one-tenth of the time constant, the sampling period is set to 3s. Finally, a total of 4,800 sets of data were collected. The variables included in the data are cold aisle temperature, supply and return air temperature, cold aisle wind speed, supply air speed, inlet temperature of cabinets 1 to 8, and load. The original sample data is as follows Fig.11 The data set is randomly divided in a ratio of 8:2, that is, 3840 groups are selected as training data and the remaining 960 groups are used as test data.

[0306] Since the variables of the collected original data have different dimensions and the numerical values ​​vary greatly, the training time of the neural network model will be too long, or even fail to converge. Therefore, the collected data is normalized and denormalized. The normalization formula and denormalization formula are: In both formulas, represents the sample data after normalization; X represents the sample data before normalization; X max , X min They correspond to the maximum and minimum values ​​of X respectively.

[0307] Determination of the number of fuzzy clusters in the prediction model

[0308] The number of fuzzy clusters (number of fuzzy rules) has a great influence on the results of TS fuzzy modeling. Therefore, before identifying the antecedent structure and consequent parameters of the prediction model, it is necessary to determine the number of fuzzy clusters. The enumeration method is used to determine the optimal number of fuzzy clusters l based on the MATLAB environment.

[0309] First, the maximum number of fuzzy clusters l is preset based on experience max =10, set l=2 and loop until l=l max , compare the impact of different l on the model performance indicators. The fitness calculation in Table 5 is shown in the objective function of FCM clustering. The performance evaluation indicator of the model is the mean absolute error (MSE), and its calculation formula is: In the formula, N represents the amount of data used in the test. represents the actual expected output of the model, y k is the predicted output of the model.

[0310] From the enumeration results in Table 5, it can be found that as the number of clusters increases, the fitness of the IBAS algorithm decreases, while the modeling time increases. When the number of fuzzy clusters is 5, the performance index MSE of the model is the lowest, and the fitness and modeling time of the IBAS algorithm are also at a low level, so the optimal number of fuzzy clusters l is selected as 5.

[0311] Table 5 The impact of different cluster numbers on training results

[0312]

[0313] Comparison of Precursor Structure Optimization Algorithms

[0314] The advantages and disadvantages of the PSO-FCM algorithm and the IBAS-FCM algorithm are compared based on the MATLAB environment, and the comparison results of the optimization performance of the two algorithms are shown in Table 6. The fitness calculation in Table 6 is shown in the objective function of FCM clustering.

[0315] The parameters of the FCM algorithm are set as follows: 1) the number of clusters is 5; 2) the total number of iterations is 300.

[0316] The parameters of the PSO algorithm are set as follows: 1) inertia factor 0.75; 2) individual experience learning factor 0.5; 3) social experience learning factor 1.5.

[0317] The parameters of the IBAS algorithm are set as follows: 1) search step size 0.1; 2) decay factor 0.99; 3) step size factor 1; 4) crossover probability 0.3; 5) crossover factor 0.02; 6) mutation probability 0.3.

[0318] Table 6 Comparison of optimization performance of the two algorithms

[0319]

[0320] It can be seen from Table 6 that compared with the PSO-FCM algorithm, the IBAS-FCM algorithm has a smaller fitness and takes less time, which means that the algorithm can shorten the antecedent structure identification time and has a higher optimization accuracy.

[0321] Comparison of post-processing parameter identification methods

[0322] When the IBAS-FCM algorithm is used for antecedent structure identification, the prediction model performance of EKF, UKF, and CKF identification methods is compared based on the MATLAB environment. The comparison results are shown in Table 7. The model performance evaluation index is MSE.

[0323] Table 7 Comparison of model performance under different consequent parameter identification methods

[0324]

[0325] It can be seen from Table 7 that the MSE and relative error of CKF are lower than those of EKF and UKF. Therefore, using the CKF method to identify the consequent parameters of the TS fuzzy model can make the model have higher prediction accuracy.

[0326] Predictive model performance analysis

[0327] Based on the above-designed computer room temperature prediction model structure, the number of fuzzy rules and the identification results of the model's antecedents and aftermath, the improved TS fuzzy model is trained using the divided training data set, and the performance of the improved TS fuzzy model is tested using the divided test data set. Fig.12 The test results of the room temperature prediction model are shown in the figure. The horizontal axis represents 960 sets of test data. Fig.12 It can be seen that the predicted value output by the prediction model can track the actual value well, and the relative error between the predicted value and the actual value is between -1% and 1%. In addition, the performance index MSE of the prediction model is 6.7669e-05, indicating that the established room temperature prediction model can effectively represent the dynamic response and transition process of the room air conditioning area temperature, and has engineering practical value.

[0328] Control strategy experiment

[0329] At present, most cooling unit systems in data centers are controlled by PID method. Therefore, simulation experiments are carried out to compare the neural network predictive control with fixed weight coefficient (hereinafter referred to as "fixed weight predictive control"), the neural network predictive control with the optimization target weight adaptive module (hereinafter referred to as "weight adaptive predictive control") and PID control to verify the performance of the proposed method.

[0330] Simulation experiment environment setting

[0331] For a system with inertia and hysteresis, if the prediction time domain in the predictive control is too small, it will affect the control performance. If the prediction time domain is too large, it will increase the amount of calculation to solve the optimization problem, making the dynamic response slower. Generally speaking, the prediction time domain should cover the entire transient range of the controlled object. In view of the small area of ​​the machine room, the air outlet of the air conditioner is close to the cold channel, and based on the principle that the control cycle should be greater than or equal to the sampling cycle, the simulation selects the prediction and control cycle as 6s, and the prediction time domain M is 10 prediction steps, that is, 60s, to cover the entire transient range of the controlled object; the number of simulation steps is set to 150, which is equivalent to 900s of actual operation.

[0332] The initial variable values ​​of the simulation are shown in Table 8. The data center design specification (GB50174) requires the cold channel temperature or the temperature near the cabinet entrance to be 18℃~27℃, so the setting value of the maximum cabinet entrance temperature T set The temperature is set to 27℃ to save the energy consumption of the cooling system as much as possible. In addition, during the simulation experiment, the load changes and chilled water flow changes collected from the actual data center cooling system are introduced as system disturbances.

[0333] Table 8 Initial variable values ​​of simulation study

[0334]

[0335] Controller parameter settings

[0336] The parameter settings of the neural network controller for fixed-weight predictive control and weight-adaptive predictive control are shown in Table 9.

[0337] Table 9 Neural network controller parameters

[0338]

[0339] At present, the cooling unit system of the data center usually adopts a single-loop system structure, such as PID control. However, PID control is a post-effect control, which is sensitive to disturbances. Therefore, it is not compared with this method, but with the improved PID control, in which the air supply volume control loop adopts PID control and the air supply temperature control loop adopts cascade PID control.

[0340] like Fig.13 As shown, the PID control structure designed in this embodiment is implemented by two PID controllers and one PI controller, one of which outputs a fan control signal, and its setting parameter is K p1 , T i1 , T d1 Another PID controller outputs the reference input of the inner loop controller of the supply air temperature control loop, and its tuning parameter is K p2 , T i2 , Td2 The PI controller in the inner loop of the air supply temperature control loop outputs the water valve control signal, and its setting parameter is K p , T i It is worth mentioning that in order to facilitate the comparison of the control performance of each controller in the outer loop, the inner loop PI controller of the supply air temperature control loop of the predictive control and PID control is the same.

[0341] This embodiment uses a particle swarm optimization (PSO) algorithm to adjust the parameters of the inner loop PI controller of the supply air temperature control loop in the predictive control and the parameters of each controller in the PID control. The adjustment steps are as follows:

[0342] 1) Randomly generate N initial particles and set the initial parameters of each initial particle, including the particle's position, speed, individual optimal position, and global optimal position;

[0343] 2) According to the initial particle swarm position, calculate the fitness value of each particle. The fitness function selected here is: In the formula, k m Indicates the number of simulation steps of control; e f [k] represents the deviation between the actual value of the chilled water flow and its set value.

[0344] 3) Compare the calculated fitness value of each particle with the individual optimal value and the global optimal value, update the individual optimal position and the global optimal position, and update the individual optimal value and the global optimal value at the same time;

[0345] 4) Update the speed of each particle. The speed update formula is: In the formula, represents the velocity of the i-th particle at the d-th iteration; represents the speed of the ith particle in the d-1th iteration; w represents the inertia factor; c1 represents the individual experience learning factor; c2 represents the social experience learning factor; rand() is a random value in the range of [0,1]; represents the position of the i-th particle in the d-1-th iteration; represents the individual optimal position of the i-th particle; represents the global optimal position.

[0346] 5) Update the position of each particle. The position update formula is:

[0347] 6) Update the fitness of each particle according to the fitness function, and update the individual optimal position and the global optimal position;

[0348] 7) Iteration number d = d + 1, repeat steps 4) to 6) until the maximum number of iterations is reached, and finally the global optimal position is is the best tuning parameter.

[0349] In summary, the parameter tuning results of each controller are shown in Table 10.

[0350] Table 10 Parameters of neural network predictive control and PID control

[0351]

[0352] In the fixed-weight predictive control, in order to reduce the temperature error and fan energy consumption, combined with the experience of weight coefficient adjustment, the weight coefficients of each item were obtained through trial and error: α = 0.4, β = 0.2, γ = 0.4.

[0353] Comparison of experimental results between predictive control and PID control

[0354] The various parameter settings of the simulation comparison experiment are as described above.

[0355] Comparison of control effects under different control methods Fig.14 As shown in the first picture. Fig.14 It can be seen that under the conditions of the same system disturbance, the neural network predictive control can achieve the control target in a shorter time than the PID control; the maximum cabinet inlet temperature of the PID control fluctuates more significantly near the set value than the neural network predictive control, and the overshoot is larger, about 2.6%, indicating that the neural network predictive control method has better control effect; after adding the weight adaptive module, the neural network predictive control fluctuates less near the set value, and the control effect is further improved.

[0356] Comparison of chilled water flow setting values ​​under different control methods Fig.14 As shown in the second figure in the figure. It can be seen that compared with PID control, when the system disturbance continues to occur, the range of change of the chilled water flow set value of the neural network predictive control is smaller, so that the water valve action of the inner loop control loop changes less, effectively reducing the energy consumption of the refrigeration end water pump; after adding the weight adaptive module, the chilled water flow set value increases slightly compared to before adding, indicating that when the energy consumption generated by the water valve itself is low, its weight in the optimization objective function decreases, and more weight is allocated to the fan control signal item.

[0357] Comparison of fan control signals under different control methods Fig.14As shown in the third figure in the figure. It can be seen that the fan control signal of the neural network predictive control is smaller than the fan control signal of the PID control most of the time, and the change is smoother and the change range is smaller, indicating that the neural network predictive control can keep the fan control signal relatively stable and avoid frequent fan operation to extend its service life; after adding the weight adaptive module, the fan control signal is slightly reduced compared with before adding, indicating that when the energy consumption of the cooling unit system is mainly generated by the variable frequency fan, the overall performance of the system is more focused on saving fan energy consumption when the deviation e is small.

[0358] The weight adjustment curves in the weight adaptive predictive control are as follows: Fig.14 As shown in the fourth figure of the figure, it can be seen that the weight-adaptive predictive control can adjust α, β, and γ in real time according to the deviation e and the difference ΔT, effectively suppressing the impact of the actual system disturbance on the maximum temperature of the cabinet inlet and improving the overall performance of the system.

[0359] As mentioned above, the energy consumption of the cooling system is mainly generated by the variable frequency fan. The fan energy consumption of different control methods is as follows.

[0360] Table 11 Comparison of fan energy consumption under different control methods

[0361]

[0362] It can be seen from Table 11 that compared with PID control, the fixed-weight predictive control can save about 44.74% ((0.2264-0.1251) / 0.2264×100%) of fan energy consumption, while PID control cannot optimize the overall performance of the system and cannot meet the requirements of energy saving; after adding the weight adaptive module, the fan energy consumption is further saved by about 16.31% ((0.1251-0.1047) / 0.1251×100%).

[0363] In summary, under the same conditions, the prediction control strategy proposed in this embodiment can effectively improve the system dynamic performance and has high energy saving potential due to feedback, multi-step prediction and rolling optimization. The addition of the weight adaptive module enables the prediction control to take into account both adaptive performance and energy saving in the continuous adjustment process, which not only ensures the adaptive performance of the system, but also further saves the fan energy consumption, thereby improving the overall control effect of the system.

[0364] Analysis of control effect under different initial values

[0365] In order to prove the stability and robustness of the strategy proposed in this embodiment, two groups of simulation comparisons were carried out on the basis of weight adaptive predictive control. Three different sets of initial variable values ​​were selected for comparison in each group. The specific variable values ​​are shown in Table 12 and Table 13. The changes of the maximum temperature of the cabinet inlet under different initial variable values ​​are shown in Table 12 and Table 13. Fig.15 As shown, it can be seen that the control strategy proposed in this embodiment has achieved a relatively stable convergence effect under different initial variable values ​​and has good robustness.

[0366] Table 12 The first set of initial variable values

[0367]

[0368]

[0369] Table 13 The second set of initial variable values

[0370]

[0371] The cooling unit system of a data center has a complex structure, with nonlinear, large hysteresis and strong coupling characteristics. In addition, since there are multiple cooling units in the data center room, the state change of a unit will cause the pressure of the chilled water pipe to change, which is easy to cause the air supply temperature to fluctuate for a long time, which brings difficulties to the control system design. In addition, the existing nonlinear optimization algorithm has a large amount of calculation and is difficult to implement in engineering applications. Therefore, in order to meet the safe operation of the data room, save fan energy consumption, and reduce the chilled water flow, and to reduce the energy consumption of the refrigeration end water pump as the system optimization goal, a model predictive control strategy for the cooling unit system of a data center is proposed. Firstly, in order to solve the problem that the pressure fluctuation of chilled water pipeline easily leads to long-term fluctuation of supply air temperature, the supply air volume control loop and supply air temperature control loop are designed. The supply air volume control loop adopts direct predictive control structure, and the supply air temperature control loop adopts cascade predictive control structure. The outer loop adopts predictive control to achieve multi-objective optimization and ensure the control accuracy of the system. The inner loop adopts PI control based on PSO optimization to suppress the frequent disturbance of chilled water flow change and ensure the rapid response of the system. Secondly, the mathematical model of the cooling unit components is constructed based on the mechanism modeling method, and the improved TS fuzzy room temperature prediction model is designed (to improve the accuracy and portability of the prediction model). The proposed TS fuzzy room temperature prediction model is adopted. IBAS-FCM and CKF algorithms are used to identify the antecedent structure and consequent parameters of the TS fuzzy model, respectively, to improve the accuracy and portability of the model; in order to solve the problem of difficulty in nonlinear optimization, a neural network predictive control rolling optimization algorithm is proposed, which uses a neural network as a feedback optimization controller, takes the system optimization objective function as the controller optimization performance index, and uses the variational method and stochastic gradient descent method to dynamically optimize the controller weights and thresholds to obtain the optimal control quantity sequence; in order to solve the problem of conflicts among sub-objectives in the multi-objective optimization process, an optimization objective function weight adaptive module is designed based on fuzzy logic to dynamically and adaptively adjust the weight coefficients in the system optimization objective function. For a data center in Beijing, a computer room CFD model is established using the computational fluid dynamics method to obtain system operation data, and the construction of the cooling unit component model and the TS fuzzy prediction model is completed. On this basis, a control strategy simulation experiment is conducted. The simulation experiment results show that the predictive control strategy proposed in this example has better dynamic performance than PID due to feedback, multi-step prediction and rolling optimization. Under the premise of ensuring the safe operation of the data center, the fan energy consumption is saved by about 44.74%. The weighted adaptive predictive control takes into account both adaptive performance and minimal energy consumption, further saving the fan energy consumption by about 16.31%. Moreover, a relatively stable control effect is achieved under different initial values ​​of the variables, proving that the proposed control strategy has good stability and robustness.

[0372] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A nonlinear model predictive control system for a cooling unit system in a data center, characterized by: The nonlinear model predictive control system includes an air supply volume control loop and an air supply temperature control loop; The air supply volume control loop adopts a direct predictive control structure, the control variable is the fan frequency conversion control signal, and the controlled variable is the maximum temperature of the cabinet inlet; The air supply temperature control loop adopts a cascade predictive control structure; The output control quantity of the outer loop of the cascade predictive control structure is the chilled water flow set value, and the controlled variable of the outer loop of the cascade predictive control structure is the maximum temperature of the cabinet inlet; The output control quantity of the outer loop is the reference input of the inner loop of the cascade predictive control structure; The output control quantity of the inner loop of the cascade predictive control structure is the water valve control signal, and the controlled variable is the chilled water flow rate; The outer loops of the direct prediction control structure and the cascade prediction control structure both use a prediction controller; The predictive controller performs predictive control according to a data center cooling unit model.

2. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 1, characterized in that: The inner loop of the cascade predictive control structure adopts a PI controller; The parameters of the inner loop PI controller are tuned using the PSO method.

3. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 1, characterized in that: The data center cooling unit model includes a data center cooling unit system component model; The data center cooling unit system component model is constructed by using a mechanism modeling method, and the data center cooling unit system component model includes a surface cooler component model, a water valve component model, and a fan component model; The surface cooler component model is: Difference equation on the chilled water side: Where, T wi (k) represents the chilled water temperature at the inlet of the condenser at time k; T wo (k) represents the chilled water temperature at the outlet of the condenser at time k; m w Indicates the chilled water mass flow rate; M w Indicates the quality of chilled water stored inside the surface cooler; T m represents the sampling period; c w It represents the specific heat capacity of chilled water at constant pressure; The difference equation on the air side: Where, T ai (k) represents the inlet air temperature of the cooler at time k; T ao (k) represents the outlet air temperature of the condenser at time k; ξ represents the moisture separation coefficient; m a Indicates air mass flow rate; M a Indicates the quality of air stored inside the cooler; c a represents the specific heat capacity of air at constant pressure; Q(k) in the differential equation on the chilled water side and the differential equation on the air side represents the heat exchange between the air and the chilled water. The calculation formula of Q(k) is: Q(k)=FhΔt m In the formula, F represents the heat exchange area of ​​the surface cooler; h represents the heat transfer coefficient; Δt m represents the logarithmic mean temperature difference; The water valve component model is: G=2.1×10 -10 L 7 -7.7×10 -8 L 6 +1.1×10 -5 L 5 -7.1×10 -4 L 4 +2.1×10 -2 L 3 -0.19L 2 +0.16L+0.14 In the formula, G represents the relative flow rate of chilled water; L represents the relative opening of the water valve; The fan component model is: The calculation formula of the fan's air outlet speed is: v f =v fr λ f In the formula, v f Indicates the air outlet speed of the fan; v fr Indicates the rated wind speed of the fan; f Indicates the fan control signal; the fan outlet speed calculation formula is established under the assumption that the fan control signal is 0-100%; Fan energy consumption calculation formula: Where, P represents the energy consumption of the fan; P r Indicates the energy consumption of the fan under rated conditions; a i Represents polynomial factors, where i = 0, 1, 2, 3, ….

4. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 1, characterized in that: The data center cooling unit model includes a computer room temperature prediction model; The input variables of the room temperature prediction model include the chilled water flow set value f w [k], fan frequency conversion control signal u fan [k], air supply temperature T in [k], maximum cabinet inlet temperature T max [k] and the load of each cabinet; the output variable of the room temperature prediction model is the maximum cabinet inlet temperature T at the next moment. max [k+1]; The computer room temperature prediction model is a computer room temperature TS fuzzy model; The TS fuzzy model output formula of the room temperature TS fuzzy model is: In the formula, β ik represents the applicability of s(k) to the i-th rule, y i (k+1) represents the linear output of the i-th rule; a i =[a i0 ,a i1 ,...,a in ], a i Represents the set of consequent parameters of the i-th rule.

5. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 4, characterized in that: The antecedent structure identification method of the TS fuzzy model of the room temperature is as follows: by taking the objective function of FCM as the fitness function of the BAS algorithm, taking the cluster center as the beetle position in the BAS algorithm, and continuously iterating to find the optimal beetle position, so that the clustering objective function is minimized; The preceding structure identification method comprises the following steps: 1) Initialization parameters; 2) Randomly generate the initial position of the longicorn; 3) Calculate the search direction and sensing length respectively; 4) Update the antenna position by searching the direction and sensing length; The updating formula of the antenna position is: In the formula, x t is the antenna position of the longicorn beetle at the tth iteration, is the search direction, p t is the sensing length, is the position of the left antenna at the tth iteration, is the position of the right antenna at the tth iteration; 5) Calculate the antenna position according to the membership update formula and The corresponding membership matrix is ​​used to calculate the antenna position according to the objective function of FCM clustering. and The corresponding fitness; 6) Update the position x of the beetle antenna according to the following formula: t , and update the global optimal position of the longicorn according to the fitness; In the formula, x t-1 is the antenna position of the longicorn beetle in the t-1th iteration; f s (·) represents the fitness function; sign(·) is the sign function; 7) Determine whether to perform the next iteration. If the next iteration is performed, t is incremented by 1 and jumps to step 3); 8) Obtain the globally optimal rule antecedent parameters through calculation.

6. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 5, characterized in that: In step 4) of the antecedent structure identification method, the antenna position is also optimized; The optimization operation is to perform a crossover operation or a mutation operation according to the antenna position to obtain a new antenna position, and if the new antenna position is better than the antenna position, the antenna position is replaced by the new antenna position; The crossover operation is to perform a crossover operator according to the antenna position by the following formula when the crossover condition is met: In the formula, α represents the crossover factor, is the position of the left and right antennae after arithmetic crossover, is the position of the left antenna at the tth iteration, is the position of the right antenna at the tth iteration; The mutation operation is to execute the mutation operator according to the antenna position by the following formula when the mutation condition is met: In the formula, x best represents the global optimal position of the longicorn; Gaussian(1,1) represents a Gaussian distributed random number with a mean and variance of 1. and are the positions of the left and right antennae after Gaussian mutation, is the position of the left antenna at the tth iteration, is the position of the right antenna at the tth iteration.

7. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 4, characterized in that: The subsequent parameter identification method of the TS fuzzy model of the room temperature is the CKF method; The nonlinear system used for parameter estimation in the CKF method is: a(k)=f h (a(k-1))+w(k-1) z(k)=h(a(k))+v(k) In the formula, a(k) represents the system state at time k, that is, the subsequent parameter of the TS fuzzy model of the room temperature to be identified, where 0≤k≤h; w(k-1) and v(k) represent Gaussian white noises with zero mean and covariances of Q(k) and R(k) respectively, which are independent of each other; z(k) represents the system measurement value at time k; f h (·) and h(·) represent the system equation and the measurement equation respectively. h(·) is obtained by the TS fuzzy model output formula of the room temperature TS fuzzy model. h (·) is obtained by the following formula: f h (a(k-1))=ψ(k-1)a(k-1) Where Ψ is a constant matrix, a(k-1) represents the system state at time k-1, and 0≤k-1≤h.

8. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 1, characterized in that: The prediction controller is a neural network prediction controller, and the prediction control method of the neural network prediction controller is a neural network nonlinear rolling optimization method; The input variables of the neural network prediction controller include supply air temperature, cabinet load, cabinet inlet maximum temperature, cabinet inlet maximum temperature setting value, and -1 related to the neural network threshold; The output variables of the neural network prediction controller include a chilled water flow set value and a fan frequency conversion control signal; The number of neurons in the hidden layer of the neural network predictive controller is 6; The optimization goal of the neural network nonlinear rolling optimization method is to minimize the error between the maximum temperature at the cabinet inlet and the set value, while maximally saving fan energy consumption and reducing the chilled water flow; The optimization objective function of the neural network nonlinear rolling optimization method is: Where, J(k) represents the prediction control optimization objective function; M is the prediction time domain; t1 is the starting time of the prediction time domain; T set represents the set value of the maximum temperature at the cabinet inlet; α represents the weight coefficient of the temperature tracking error term; f w [k] can reflect the change of water valve opening from the side, β represents f w [k] is the weight coefficient of the term; γ represents u fan The weight coefficient of the [k] term.

9. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 8, characterized in that: The weight coefficient of the optimization objective function is obtained by optimizing the objective weight adaptive module; The workflow of the optimization target weight adaptive module is as follows: s01: Calculate the first deviation value at time k according to the maximum temperature at the cabinet inlet at time k and the set value of the maximum temperature at the cabinet inlet, and calculate the second deviation value at time k according to the maximum temperature at the cabinet inlet at time k and the maximum temperature at the cabinet inlet at time k-1; s02: Calculate the weight coefficient α[k] of the tracking error term at time k according to the first deviation value at time k; α[k]=α[k-1]+K pf (e[k]-e[k-1]) In the formula, α[k-1] is the weight coefficient of the tracking error term at time k-1, e[k-1] is the first deviation value at time k-1, e[k] is the first deviation value at time k, and K pf represents the proportionality coefficient; s03: Determine whether the first deviation value at time k is within the domain of the first deviation value. If so, use the fuzzy system to calculate the weight coefficient β[k] of the chilled water flow set value item at time k according to the first deviation value at time k and the second deviation value at time k. Otherwise, the weight coefficient β[k] of the chilled water flow set value item at time k is the weight coefficient β[k-1] of the chilled water flow set value item at time k-1. s04: Calculate the weight coefficient γ[k] of the fan variable frequency control signal item at time k according to the weight coefficient α[k] of the tracking error item at time k and the weight coefficient β[k] of the chilled water flow set value item at time k; The input of the fuzzy system is the first deviation value and the second deviation value, and the corresponding fuzzy linguistic variables are E and EC; the output is the correction amount of the weight coefficient of the chilled water flow set value item, and the corresponding fuzzy linguistic variable is U; The domains of the correction amounts of the weight coefficients of the first deviation value, the second deviation value and the chilled water flow rate setting value item are [-0.04, 0.04], [-0.002, 0.002] and [-0.03, 0.03] respectively; The number of fuzzy linguistic values ​​of the fuzzy linguistic variables E, EC, and U are 4, 5, and 5 respectively; The fuzzy linguistic variables E and EC adopt a trapezoidal membership function, and the fuzzy linguistic variable U adopts a triangular membership function; The weight coefficient adjustment strategy of the optimization target weight adaptive module is: when the first deviation value is large, α increases, β decreases, and γ remains unchanged; when the first deviation value is small, α remains unchanged, β decreases, and γ increases; when the first deviation value is small but the maximum temperature at the cabinet inlet fluctuates around its set value, α remains unchanged, β increases, and γ decreases, α represents the weight coefficient of the tracking error term, β represents the weight coefficient of the chilled water flow set value term, and γ represents the weight coefficient of the fan variable frequency control signal term.

10. The nonlinear model predictive control system for a cooling unit system in a data center according to claim 8, characterized in that: The neural network nonlinear rolling optimization method comprises the following steps: (1) Initialize the weights of the neural network predictive controller, set the weights of the neural network predictive controller to be random numbers in the range of [-1, 1], and select the prediction time domain M and learning rate μ of the neural network predictive controller; (2) Assume that the current time k = t1, and set the expected value x at each time in the prediction time domain * [k], k = t1, ..., t1 + M-1, and the actual system output at the previous moment is used as the initial state value x[t0] of the rolling optimization at the current moment; (3) Setting the weight coefficient of the optimization objective function; (4) The state quantity of the system at time t1 is x[t1] = {T in [t1],R1[t1],R2[t1],R3[t1],R4[t1],R5[t1],R6[t1],R7[t1],R8[t1],T max [t1]}, the setting value of the maximum temperature of the cabinet inlet at time t1+1 is T set [t1+1] and -1 related to the neural network threshold act on the neural network predictive controller to obtain the control quantity u[t1]={u fan [t1],f w [t1]}; then apply x[t1] and u[t1] to the room temperature prediction model to obtain the predicted value of the state quantity at the next moment (5) The predicted value of the state quantity at time t1+i The maximum temperature setting value of the cabinet inlet at time t1+i+1 is T set [t1+i+1] and -1 related to the neural network threshold act on the neural network predictive controller to obtain the control quantity u[t1+i], where i represents the current prediction step number, i=1,2,3,…,M-1; then u[t1+i] acts on the room temperature prediction model to obtain the predicted value of the state quantity at time t1+i+1 And store all calculation results; (6) Based on the results of steps (4) and (5), calculate the Lagrange multipliers λ[k], q[k] from back to front according to the Hamiltonian canonical equation, k = t1+M-1, ..., t1+1, t1; The Hamiltonian canonical equation is: (7) adjusting the weights of the neural network predictive controller based on the stochastic gradient descent method according to q[k] calculated in step (6); (8) Repeat steps (4) to (7) and continuously adjust the weights of the neural network until the weights no longer change, that is, ΔW = 0; Apply the control quantity at this time to the cooling system of the data center to obtain the maximum cabinet inlet temperature T at the next moment. max [t1+1]; (9) Update the time, set k = t1 + 1, repeat steps (2) to (8), and perform a new round of optimization until the entire control process is completed.

Citation Information

Patent Citations

  • Nuclear power device two-loop multi-variable integrated model fuzzy predication control method

    CN101169622A

  • Method for controlling main steam temperature in thermal power plant

    CN103225799A

  • Directly-heated type heat pump system constant-temperature flow control method based on neural network

    CN106369834A

  • Tail end cooling capacity equilibrium control method

    CN107421029A

  • A method for predictive control of a data cent air conditioning system

    CN108990383A