Research On Building Energy Efficient Method Based On Parallel Reinforcement Learning

In the world,building energy consumption accounts for a relatively high proportion of the total social energy consumption,and it has been growing at a relatively rapid rate.Building energy efficiency has become the primary goal of all countries' energy policies.80% of building energy consumption is energy consumption for building operation.Therefore,an effective means to reduce building energy consumption is to use science and technology to control relevant equipment in the building.In addition,in the field of building energy efficient control,traditional control methods often have problems such as poor stability and slow convergence.With the development of artificial intelligence technology,the concept of energy conservation in smart buildings has gradually gained the attention of researchers,and research on related intelligent control methods has also become a research hotspot.Among them,the reinforcement learning method is the research focus of current smart building energy efficient methods.This paper mainly focuses on building energy efficiency issues.It mainly studies the method and framework for building energy conservation based on reinforcement learning.The specific content includes the following three parts:(1)In order to solve the problem of the slow convergence speed of the reinforcement learning control method in the field of building energy conservation control.This paper combines the multi-threading technology and the experience replay method to propose a multi-threading parallel reinforcement learning algorithm(MPRL).MPRL mainly consists of two parts.The first is based on fuzzy clustering reinforcement learning multi-threading partitioning method: the policy vector is allocated to different threads through fuzzy clustering to evaluate the policy.The second is the parallel reinforcement learning framework: the parallel running policy evaluation process and the environment interaction process.At the same time,it introduces the experience replay technology,stores the samples generated in the interaction into the sample pool,and uses the samples to update the value function repeatedly.This method can effectively speed up the algorithm learning process.Comparing MPRL with Q-Learning,Sarsa,and KCACL algorithms,they were applied to the random walk problem,the windy grid world problem,and the cart pole problem.Experimental data shows that MPRL has better convergence performance and faster learning rate.(2)Markov decision process modeling for building energy-saving control problems is proposed,and a method based on reinforcement learning adaptive control(RLAC)is proposed to solve the optimal control policy of related equipment in buildings to achieve the purpose of energy conservation.The RLAC first models the environment and reinforcement learning signals.By interacting with the environment and using Q-Learning algorithm to update the Q-value function,RLAC can quickly find the optimal Q-value function and learn the optimal control policy.The experiment of simulating the room model is compared with two traditional control methods.Experimental results show that the proposed RLAC has effective energy-saving and good convergence effect,and has batter convergence rate and robustness than the other traditional control methods.(3)Combining the Markov decision process model for building energy efficiency,the parallel reinforcement learning method is used for building energy conservation,and a method of building energy conservation based on parallel reinforcement learning is proposed.This method combines the multi-threading technology and experience replay method to propose a multi-threading parallel reinforcement learning algorithm framework.By calculating the distance between samples,we select a low similarity sample to construct a diversity sample pool.The learning process of Agent is from the diversity sample pool.Selecting sample learning can effectively avoid the waste of learning resources.This method can effectively improve the algorithm learning efficiency and accelerate the algorithm learning process.The experiment includes the comparison of the simulation room model with the Q-Learning algorithm and the classical PID control method.The results show that the proposed parallel algorithm has a certain energy-saving effect,faster learning rate and convergence speed,and has higher operation effectiveness than other control methods.