Reinforcement Learning Algorithm Based On Generative Adversarial Networks And Its Application In Building Energy Conservation

With the rapid development of China’s cities,the number of large-scale public buildings is increasing,and the problem of high energy consumption in buildings is becoming more and more prominent.Therefore,building energy conservation has become a research focus in the field of building intelligence.Energy consumption prediction is an important prerequisite for energy consumption optimization.By analyzing energy consumption prediction results,it can provide a basis for building energy conservation.Large public buildings are a complex nonlinear system,which makes energy prediction difficult,and energy consumption prediction needs to collect a large number of energy samples in different states,but the collection cost of the samples is high.Reinforcement learning is a learning method that can collect samples in interaction with the environment.By interacting with the environment to obtain samples,it can learn the mapping from environment to action.The ultimate goal is to maximize the cumulative reward and obtain the optimal strategy.Its main advantage is self-learning.This paper focuses on how to use the reinforcement learning method to study the building energy consumption prediction,and solve the problem of the lack of real energy consumption samples by generating samples that are similar to historical energy consumption generated by generative adversarial networks.And then use the Q learning algorithm in reinforcement learning for building energy consumption prediction.In order to improve the performance of the algorithm and the accuracy of energy consumption prediction,deep learning and value function approximation method are introduced to construct a deep Q network.The main content of this paper includes the following three parts:(1)Aiming at the problem of the lack of real energy consumption samples in the application of reinforcement learning to energy consumption prediction,an algorithm based on generative adversarial networks is proposed.In the early stage of training,the algorithm collects experience samples through random strategies to form a real sample pool,and uses the collected experience samples to train the generative adversarial networks.Then,use it to generate new samples to form a virtual sample pool.The training samples are selected together with the real sample pool and the virtual sample pool.Then,the algorithm is applied to the reinforcement learning problem in the OpenAI Gym simulation platform.The results show that the algorithm can effectively solve the problem of insufficient experience samples in the early stage of the reinforcement learning task.(2)Applying the algorithm proposed in the previous part to the prediction of building energy consumption,a Q-learning energy prediction algorithm based on the generative adversarial networks is proposed.The algorithm models the energy consumption prediction problem as a time series prediction problem,and then combines the Q learning algorithm to predict the building energy consumption in the future.At the same time,it introduces generative adversarial networks to generate new building energy consumption samples,and solves the problem of insufficient building energy consumption samples.Finally,based on the building energy consumption data of Baltimore Gas and Electric Power Company of the United States,the proposed algorithm is analyzed.The results show that the proposed algorithm can effectively predict the building energy consumption in the future.(3)Aiming at the problem of poor performance of Q learning algorithm applied to nonlinear energy consumption prediction,a deep Q learning energy prediction algorithm based on generative adversarial networks is proposed.The algorithm introduces a deep neural network and constructs a deep Q network to calculate the action value function,the input is the state,the output is the action value function of each action,and the value function approximation method is used to avoid the problem that the Q learning algorithm is poor in the large state space.The results show that the proposed algorithm can further improve the prediction accuracy of building energy consumption.