Research On Energy Saving Method Of HVAC For Large-scale Search Space

The building sector,on a global scale,accounts for nearly 40% of the total energy consumption of society and contributes 30% of the total CO2 emissions.Building energy consumption accounts for nearly 20.6% of the total social energy consumption in China and 19.4% of total CO2 emissions.Heating,ventilating,and air-conditioning(HVAC)is considered as the largest energy consumption system in a building,since more than 50%of the building energy consumption is from it.Controlling HVAC systems by using more efficient optimization algorithms has advantages in terms of efficiency,sustainability and economic benefits over replacing HVAC system equipment.Various optimization methods proposed for HVAC systems in recent decades,such as traditional mathematical optimization method,heuristic method,machine learning method etc.Among them,machine learning method is not limited by the exact objective function compared with traditional mathematical optimization methods,and it is more robust with stable convergence results than heuristic method.Therefore,machine learning method,especially deep reinforcement learning method,has great research value in energy saving optimization.This paper focuses on how to use heuristic method and deep reinforcement learning method to study the energy-saving optimization strategy of large-scale central HVAC.Based on the constructed HVAC simulation system,the influence of the total energy consumption of the HVAC subsystems is analyzed,and deep reinforcement learning with heuristic method is used to find the optimal set values of each HVAC subsystems.According to the optimal value,the cooling water temperature,the primary chilled water temperature,the secondary chilled water temperature and the air supply temperature in the HVAC system are set.As the state space composed of the set values of each subsystem of HVAC can reach tens of thousands,the traditional optimization method cannot obtain the optimal set value combinations in time,and the traditional deep reinforcement learning method is limited by the problem of "dimension disaster" and cannot converge in a limited period of time.In this paper,two improved HVAC optimization algorithms are proposed through sensitivity analysis,function approximation,network pruning,which greatly reduce energy consumption of HVAC.The main contents include the following three parts:(1)A heuristic optimization algorithm based on sensitivity analysis is proposed to address the problem that traditional optimization methods cannot get the optimal set value in time in the face of a large-scale search space.The heuristic method does not need to traverse all state spaces,so a better set value can be obtained in a reasonable time.Through sensitivity analysis method for HVAC subsystems to quantitative analysis of the influence of the total energy consumption,and based on the results of sensitivity analysis,HVAC subsystems can be divided into sensitivity and non-sensitivity factors.The set values of the subsystems in the sensitive factors are first optimized in combination,and on the basis of this optimization result,the set values of the subsystems in the non-sensitive factors are then optimized in combination.The proposed combinatorial optimization algorithm based on sensitivity analysis greatly reduces the size of the traversal space,thus solving the problem that the traditional optimization method cannot get the optimization results in time in the face of large-scale search space,and the simulation results show that the combination optimization method based on sensitivity analysis has remarkable energy-saving effect.(2)Aiming at the problem that the traditional deep reinforcement learning method cannot converge in a limited time under the large-scale state space,an optimization algorithm of multi-step actor critic is proposed.By pruning the traditional deep reinforcement learning network,the algorithm greatly reduces the convergence time and the data required for training.A concept of basic actions is proposed to reduce the traditional deep reinforcement learning network by using basic actions instead of the full set of actions,and by selecting different basic actions to form multiple optimization networks.Further,by combining with the search tree to form a multi-step actor module,the network structure with the best energy-saving effect is finally selected,and the simulation results show that the multi-step actor optimization algorithm achieves an energy-saving effect similar to that of traditional deep reinforcement learning.(3)In view of the performance losses existing in the multi-step actor module based on pruning deep reinforcement learning network and search tree in the second part,the performance is further optimized by adding a critic module into the framework of traditional deep reinforcement learning methods.The critic module optimizes the set value again by KNN and value function to further reduce the energy consumption of HVAC.Finally,the simulation environment is constructed to verify that the multi-step actor critic optimization algorithm can achieve the optimal performance and the lowest energy consumption compared with the traditional deep reinforcement learning and many traditional optimization algorithms.