Modeling And Predictive Control For Air Conditioning Terminal Systems In Data Center

In recent years,with the rise of cloud computing,5G commercialization and meta-universe,a large number of data centers have been invested,built and developed,and their excessive energy consumption has become an increasingly important issue.Data center air conditioning systems need to run non-stop throughout the year,and their energy consumption cannot be ignored.The data center air conditioning system is divided into two parts: the cooling station system and the terminal system,of which the air conditioning terminal system is responsible for providing cold air to the computer room to ensure reliable operation of IT equipment.According to statistics,the energy consumption of air conditioning systems accounts for about 29% of the total energy consumption of data centers,while the terminal energy consumption accounts for about 71% of the energy consumption of air conditioning systems.Part of the reason for the high energy consumption of data center air conditioning terminal systems is that the system has non-linearity,inertia,time lag,and is easily disturbed by chilled water flow from other branches,making it difficult for PID to achieve overall system energy efficiency improvement;another part of the reason is that to avoid the risk of overheating in the racks,conservative users usually increase the air supply volume and reduce the air supply temperature,but this approach tends to lead to excessive cooling,resulting in additional energy waste.In view of the fact that predictive control can improve the dynamic characteristics of the object,the predictive model and rolling optimization mechanism can be used to realize the system adaptive optimization search,while the cascade control is suitable for the system with frequent disturbances,this thesis combines cascade control with predictive control to investigate the cascade predictive control strategy of the data center air conditioning terminal system by making the rack inlet temperature meet the cooling demand and save energy as the system optimization goal.The basis of predictive control is predictive modeling,but the complex structure and numerous devices of data center air conditioning terminal systems make it difficult to model the mechanism,so this thesis uses a data-driven approach to construct the system model.Since it is difficult to obtain operational data in actual engineering and difficult to reflect the complete dynamic characteristics of the system,this thesis uses the numerical simulation method of computational fluid dynamics(CFD)to establish a CFD model of the computer room,constructing a 3D geometric model,a mesh model,boundary conditions,etc.Considering that predictive control requires model feedback correction during rolling optimization and that model portability plays an important role in improving the efficiency of predictive control studies,this thesis investigates the T-S fuzzy model construction method for data center air conditioning terminal systems.In order to improve the recognition accuracy and the online correction performance of T-S fuzzy model,this thesis proposes the improved T-S fuzzy model,firstly,the influencing factors of the rack inlet temperature are analyzed and the model structure is determined;then,in order to avoid the fuzzy C-mean clustering algorithm(FCM)from falling into local optimum and to improve the efficiency of the search,the FCM with improved beetle antennae search(IBAS-FCM)algorithm is proposed to achieve the structural identification of model by using arithmetic crossover operator and Gaussian variational operator;cubature kalman filter(CKF)algorithm is adopted to identify the consequent parameters.The experimental results show that the improved T-S fuzzy model can shorten the time of structure identification and the model prediction accuracy is higher compared with the traditional model,and the relative error of the maximum rack inlet temperature is 1.2%.To test the portability of the model,this thesis uses the data collected from another CFD model of the computer room to calibrate the model online,and after 1160 iterations,the new model can converge and the model accuracy can reach more than 95%.In order to overcome the shortcomings of traditional PID control,this thesis proposes a neural network cascade predictive control strategy.The outer loop adopts predictive control to improve the dynamic response quality and save energy;the inner loop adopts PID control based on PSO optimization,whose reference input is the output of the outer loop controller,to suppress the effect of frequent fluctuations of chilled water flow on the rack inlet temperature.Due to the nonlinear characteristics of data center air conditioning terminal,the traditional nonlinear optimization algorithm requires extensive computational cost and storage space for online optimization.For this reason,this thesis proposes a neural network predictive control rolling optimization algorithm,which uses the neural network(NN)as a feedback optimization controller,takes the system optimization objective function as the performance index of the NN.By combining the Euler-Lagrange and stochastic gradient descent methods,the weights(with thresholds)of the NN are optimized,which is easy to implement in engineering and can be applied to the rolling optimization of the outer-loop predictive control.In addition,the weight coefficients in the predictive control optimization objectives have a certain influence on the system performance.In order to improve the overall performance of the system and resolve the contradictory relationship between more than two optimization objectives,this thesis designs a weight coefficient adaptive module and uses fuzzy logic to dynamically adjust the weight coefficients in the optimization objective function.The experimental results show that the proposed control strategy has better tracking performance and robustness compared with the cascade PID control,and the fan energy consumption is saved by about 36.59%;after the adaptive adjustment of the weight coefficients of the optimization objective function,the fan energy efficiency is improved by about 15.67% again.