Energy-Saving Optimization Control Of Data Center Refrigeration And Waste Heat Recovery System

Driven by the wave of new infrastructure,the rapid development of the digital economy has given rise to a great demand for data center construction.While data centers offer significant development opportunities,their high energy consumption remains a longstanding issue.Due to the need for uninterrupted cooling throughout the year in the data center,the energy consumption of its refrigeration system is second only to IT equipment.Therefore,optimizing control strategies for energy-saving in refrigeration systems has become a key research direction for reducing energy consumption in data centers.However,the data center refrigeration system belongs to a multi-loop complex system with characteristics such as nonlinearity,large inertia,and strong coupling.Currently,the traditional PID control commonly used in data centers can only achieve aftereffect control for single variables,which is challenging to overcome the impact of system lag characteristics and cannot achieve overall energy efficiency improvement of the data center refrigeration system.On the other hand,data centers generate a large amount of heat during operation,and recycling waste heat for heating or hot water systems can improve energy efficiency.However,the waste heat recovery system is affected by both the heating load and data room load,making the design of the control system more complicated.In response to the above issues,this paper proposes a model predictive control strategy for data center cooling systems in winter conditions,based on the full utilization of natural cold sources,to overcome the influence of system hysteresis characteristics and to improve refrigeration system energy efficiency;Furthermore,a control strategy for data center waste heat recovery system based on reinforcement learning is proposed to improve system dynamic performance and improve system operational efficiency.Firstly,a neural network was used to construct a prediction model for the refrigeration system.Based on the equipment information and operating mechanism of the research object,a data center refrigeration system model was built on the TRNSYS platform under winter conditions.By changing the load and the set value of the supply air temperature,the system’s operating status was changed to achieve data collection and capture the complete dynamic characteristics of the object.Then,the collected data was utilized,Bayesian regularization algorithm is used to train the neural network to improve the generalization ability of the model.On this basis,a data center refrigeration system model predictive control strategy is proposed to achieve the goal of safe operation of the data center and improving system energy efficiency.In order to solve the problem that the traditional nonlinear optimization algorithm has a large amount of computation and memory space,which is challenging to be realized in engineering,a three-layer feedforward neural network is used as the optimal feedback controller,and the system optimization objective function is used as the optimization performance index of the controller.Combining the Euler Lagrange method and the gradient descent method,the weight(threshold)value of the controller is optimized online to obtain the optimal control quantity sequence.In addition,this paper designs a water source heat pump to achieve waste heat recovery in the data center,which is used to provide a heat source for office buildings adjacent to the data center.Firstly,based on physical principles and system working principles,a dynamic model of the data center waste heat recovery system was constructed using Simscape Fluids.On this basis,a reinforcement learning control strategy is proposed to maintain indoor thermal comfort in buildings under constantly changing user-side thermal load and data room load.In response to the control objectives,the input and output variables of the reinforcement learning controller,as well as the structure of the policy and critic networks,were designed.A "curriculum learning" training strategy was introduced,and a stable and convergent reinforcement learning agent was obtained using a gradually increasing task complexity.The experimental results indicate that the predictive control strategy proposed in this paper for the data center refrigeration system model can improve the dynamic performance of the system and overcome the influence of uncertain factors such as inertia,lag,and interference.Compared with PID control,the reinforcement learning control strategy proposed in this paper can effectively improve the dynamic performance of the data center waste heat recovery system under constantly changing loads,reducing indoor temperature overshoot by 17.54% and regulating time by 51.06%.