Research On Real-time Home Demand Response Strategy Based On Deep Reinforcement Learning

The adverse impact of global climate change on human production and life has become more and more prominent,and has become one of the most severe challenges faced by human society.To cope with global climate change,China has proposed the goal of achieving carbon peak by 2030 and carbon neutrality by 2060.Building energy consumption accounts for about40% of global energy consumption.Home demand response is an important way for residential consumers to participate in energy conservation and emission reduction.However,it is difficult for consumers to optimize the operation of home electrical equipment due to the uncertainty of renewable energy and consumer behavior,as well as the large number of controllable devices with complex characteristics.In recent years,various machine learning technologies such as deep learning and reinforcement learning have provided effective solutions to practical engineering problems in various industries.Based on this,this thesis proposes a real-time home demand response scheduling strategy based on deep reinforcement learning,and the specific arrangement is as follows:In this thesis,a multi-step generation-load state prediction method for home microgrid with heterogeneous feature input is studied.Firstly,the home microgrid model is established according to the home demand response problem,and the functions and characteristics of various electrical equipment in the demand response scheduling scenario are analyzed.Then,an optimization model is constructed to integrate the electricity purchase cost and electricity consumption satisfaction.By analyzing the characteristics of uncertain factors in the microgrid model,the photovoltaic power and uncontrollable load power are predicted respectively,and they are incorporated into the reinforcement learning framework,so as to improve the perception and prediction ability of decision making.This thesis studies the basic strategy of home demand response scheduling based on imitation learning.In this study,the optimal scheduling problem is transformed into a Markov decision process,and the environment state is extended based on the state prediction information to form a home microgrid training environment.Aiming at the problem that the control action space dimension is too large due to the variety of controllable devices in the home demand response scenario,this thesis proposes a reduction method of action space,which effectively improves the learning efficiency and effect of the algorithm.In order to enhance the exploration ability and learning efficiency of agents in complex demand response scenarios,imitation learning based on expert rules is used to generate the basic strategy of reinforcement learning.In this thesis,an optimal scheduling strategy for home demand response based on deep reinforcement learning is studied.To improve the interactive ability and the actual operation results of reinforcement learning,a backup control strategy is proposed to assist agent decision making based on the actual requirements of consumer participation in demand response,which is involved in the training and decision process of the proximal policy optimization algorithm,and finally forms the scheduler core of the home energy management system.For each research,detailed simulation analysis is carried out with the actual home residential data as an example,and experimental comparison with other reinforcement learning algorithms verifies the effectiveness of the proposed method.